The immediate dispute revives a familiar argument. Brussels says it is policing anticompetitive conduct under the Digital Markets Act (DMA). Washington says Europe is taxing American companies. After two years of enforcement, the second account has become much harder to dismiss—which makes the Trump administration’s escalation worth taking seriously.  

Last week, the European Commission fined Google €890 million under the Digital Markets Act. It imposed €460 million for favoring Google’s own shopping, hotel, transportation, and sports results in Google Search, and €430 million for restricting developers from directing Google Play users to cheaper purchasing options. The penalty was Google’s first under the law and the largest imposed under it to date. 

The next day, Trump responded on Truth Social. He announced that the United States would “immediately” open a Section 301 investigation into what he called Europe’s “robbing” of American companies. The United States, he wrote, is not Europe’s “PIGGYBANK.” He predicted that the penalties would be “entirely reversed” and followed by a substantial tariff. 

Some of Trump’s particulars do not survive contact with the record. The $15 billion attributed to Apple appears to include the Commission’s €13 billion Irish state-aid recovery order in a tally of “fines,” though it was not a fine. Nor can a Section 301 investigation “reverse” a legally binding Commission decision. Only the European Union courts can do that. 

But the faulty arithmetic distracts from the two questions that matter: Is the underlying grievance well-founded? And, if so, is trade retaliation a defensible response? 

On the first question, the evidence has become considerably harder to dismiss. On the second, the answer is probably yes—but for reasons rooted less in trade policy than in the political economy of European regulation. 

Europe’s Industrial Policy in Antitrust Clothing

When Geoffrey Manne and I examined European Union competition enforcement in 2019, we found that U.S. firms paid vastly larger fines than European ones. Still, we concluded that the disparity largely reflected the sectors the European Commission targeted and its sales-based method for calculating penalties—not deliberate discrimination. After the DMA’s enactment and enforcement, that conclusion is harder to sustain.

The design problem is familiar. The DMA’s thresholds capture firms based on aggregate size, not demonstrated market power. Applying those arbitrary metrics, the Commission has designated seven gatekeepers. Five are American, while only one (Booking.com) is European. That imbalance is no accident. When lawmakers adopted the DMA, Andreas Schwad—member of the European Parliament for Germany—said it should focus on “the top five” companies while avoiding European rivals.  

Enforcement has made the pattern starker. Every DMA noncompliance penalty so far—€500 million against Apple, €200 million against Meta, and now €890 million against Google—has fallen on an American company. 

The Commission’s specification decisions are even more revealing than the fines. On July 16, it told Google precisely how to satisfy two DMA obligations. Google must open 11 Android system-level features to rival artificial intelligence (AI) assistants. Beginning in January 2027, it must also share anonymized Search-ranking, query, click, and view data with competing search engines and AI chatbots on fair, reasonable, and nondiscriminatory (FRAND) terms. 

These requirements go far beyond ordinary prohibitions on anticompetitive conduct. They are product-design mandates that de facto transfer parts of Google’s property to its competitors. 

Nor is the Commission coy about the destination. In its first DMA review, it candidly conceded: 

The DMA contributes to the EU’s competitiveness, innovation and technological sovereignty objectives by addressing structural bottlenecks in CPSs that may hinder market entry and scaling. By promoting contestable and fair digital markets, it complements efforts to close Europe’s scale-up gap in line with the EU Startup and Scale-up Strategy. It facilitates competition through obligations on interoperability, data portability, access to business user data, etc. and through restrictions on self-preferencing… [I]t forms part of a coherent policy framework aimed at strengthening Europe’s innovation capacity and reducing strategic dependencies.

Put plainly, the objective is digital sovereignty. The Commission is trying to bolster European rivals by redesigning—and partly expropriating—American technology companies. It is therefore no surprise that the same review identified cloud computing and AI as the next frontiers of DMA enforcement. The European Union’s June Technological Sovereignty Package, with its cloud-eligibility tiers and procurement preferences, pursues the same goal. 

The Commission’s own documents thus frame the DMA as an industrial-policy instrument. It aims to reshape the internet according to European policymakers’ preferences and for the benefit of the firms they want to succeed. 

To be fair, the DMA also binds Booking, and several of its loudest beneficiaries—including Epic and, increasingly, OpenAI—are American. The point is not that the law amounts to naked protectionism. It is that a facially neutral regime, administered by an institution facing the incentives described below, has produced a systematically one-sided result. 

The costs also extend beyond U.S. shareholders and European consumers. Compliance consumes one of the scarcest resources at frontier technology companies: senior engineers’ time. Meta reports devoting roughly 600,000 engineering hours to DMA compliance. Google says it assigned about 3,000 engineers full time for two years to comply with a single article. The Commission’s impact assessment, by contrast, projected annual compliance costs of about €1.41 million per platform—a figure that now looks almost quaint. 

Every hour spent rebuilding a codebase for Brussels is an hour not spent building something else. Google’s response to last week’s decision made the engineering and consumer costs explicit. Compliance, the company says, will require removing live hotel, flight, and restaurant prices and availability from European search results, while rolling back Play Store safeguards. American users may not pay the fines directly, but they still pay for the engineering talent diverted from improving the products they use. 

Making Brussels Feel the Cost

The deeper problem is that European policymakers pay almost no political price for enforcing the DMA more aggressively. The benefits of DMA enforcement flow to organized, well-represented constituencies: firms seeking access to rivals’ platforms, publishers, telecom incumbents, and would-be national champions. The costs are diffuse, delayed, and largely shifted onto foreign shareholders and 450 million consumers across 27 member states, who experience product degradation simply as products getting “worse.” 

The fiscal incentives point in the same direction. Fines flow into the general EU budget and reduce member states’ gross-national-income contributions. That gives governments revenue without the political unpleasantness of raising taxes. 

The contrast is telling. When the Commission blocked the Siemens-Alstom merger, Paris and Berlin issued a joint manifesto—and successfully demanded changes to the merger rules. When the Commission fines Apple, nobody blocks a motorway. 

The result is a policy that is nearly free at the point of production. European officials intend to keep it that way. They describe the DMA as settled law and refuse to treat it as a bargaining chip in trade talks, despite the August 2025 joint framework in which the EU and United States agreed to address unjustified digital-trade barriers. 

If the force driving enforcement is a public-choice problem—policymakers can impose the costs because almost nobody at home feels them directly—then the only response likely to work is one that creates a European constituency with something to lose. 

That is what targeted tariffs can do, which is why the choice of target matters far more than the headline rate. French and Italian wine, agricultural products, German automobiles, and pharmaceuticals are obvious candidates. These sectors have the organizational muscle to pressure national capitals, which can then pressure Brussels. The EU is already lobbying Washington for tariff relief on wine, spirits, olive oil, and cheese. That tells us where the pressure points are. 

Tariffs are not costless, of course, and escalation could trigger the EU’s Anti-Coercion Instrument, which allows Brussels to retaliate against foreign economic pressure with tariffs, procurement restrictions, and other limits on market access. But the relevant comparison is not tariffs against some frictionless alternative. It is tariffs against the status quo, in which the political cost to European policymakers of discriminatory regulation is roughly zero.

Now imagine the shoe on the other foot. Would the European Union sit quietly while its most successful companies were fined billions abroad under rules calibrated to capture almost exclusively them? It created the Anti-Coercion Instrument in response to considerably less. 

Does the DMA Pass the Section 301 Test?

Section 301(b) of the Trade Act of 1974 permits action against a foreign act, policy, or practice that is “unreasonable or discriminatory” and “burdens or restricts” U.S. commerce. The 1984 amendments define an unreasonable practice as one that is unfair and inequitable, even if it violates no international legal obligation, and a discriminatory practice as one that denies national or most-favored-nation treatment. The statute also expressly contemplates foreign industrial targeting. 

Because the DMA is facially neutral, de jure discrimination will be difficult to prove. The de facto case is much stronger. Designations, fines, and specification decisions have fallen disproportionately—and, in enforcement cases, exclusively—on American firms. Compliance costs exceed the Commission’s estimates by orders of magnitude, while total losses for U.S. companies may reach €114 billion. 

The harder question is credibility. As of this writing, the Office of the U.S. Trade Representative’s (USTR) public docket lists no DMA-related investigation. It is already handling 18 investigations, from forced labor in China and elsewhere to German pharmaceutical pricing. Tariffs arising from the forced-labor investigation also face a challenge before the U.S. Court of International Trade. Whether USTR has the capacity to complete another major Section 301 investigation is therefore far from clear. 

Europe, meanwhile, treats Section 301 as unilateral coercion, which raises the domestic political cost of yielding to it. Turning that confrontation into a tractable dispute will require a narrow, carefully documented case—not another press release with capital letters. 

The DMA has become an instrument of industrial policy whose costs fall overwhelmingly on a small group of foreign firms and consumers around the world, many of whom have no political voice in Europe. That is precisely the kind of problem trade law’s nondiscrimination principle exists to address. 

The goal should remain a negotiated settlement, and success should be measured by whether Europe changes how it applies the DMA—not by how much tariff revenue Washington collects. Trump’s rhetoric may be blunt, but Brussels may not reconsider the cost of DMA enforcement until French wine, German cars, or Italian cheese starts picking up the tab.

The post From Google Fines to French Wine: The Trade Case Against the DMA appeared first on Truth on the Market.

Ireland assumed the rotating presidency of the Council of the European Union from Cyprus for the second half of the year and declined to recycle Cyprus’ flawed draft. Instead, it reopened negotiations over the European Commission’s proposal on pseudonymization, cookie consent, and the use of data for artificial intelligence (AI). 

Serious reform will still face resistance in the European Parliament. A joint draft report from its industry and civil-liberties committees leaves the Commission’s most contested GDPR proposals untouched for now. But the co-rapporteurs’ public statements—and the flood of amendments filed since—show that the legislation remains very much in play. 

The European Data Protection Board (EDPB) has also weighed in. Although it does not make EU law, it exerts considerable influence over national governments. Its new guidelines distinguishing personal from anonymous data accept, for the first time, the Commission’s central point: whether data is personal should depend on the entity holding or using it, not on whether anyone, anywhere, could identify the person concerned. 

The guidelines also show why procedural reform may matter even more. As usual, the EDPB could not quite bring itself to offer guidance useful for much beyond increasing lawyers’ billable hours. 

The Reform Fight Takes Shape

The reform effort is now moving on four fronts, with the Council showing unexpected signs of ambition, Parliament preparing for a fight, the EDPB conceding a key point while muddying the details, and a separate AI measure quietly becoming law. 

The Council Hits Reset

The Cypriot presidency of the Council of the European Union tried to secure agreement among national governments on a common position that could have left the GDPR worse off than no reform at all. One week before its term ended, Cyprus circulated another compromise draft, following the version I discussed previously. 

It failed. Several governments reportedly objected, including Denmark, Germany, Italy, Poland, and Sweden. The precise reasons remain unclear, but reports suggest that these countries thought the draft did too little to simplify the GDPR. If so, there may still be hope for the reform process. 

Ireland, which now holds the rotating Council presidency, chose not to carry over the Cypriot text. Instead, it sent national governments a questionnaire reopening debate over cookie exemptions, pseudonymization, AI data processing, and reductions in compliance burdens. 

Parliament Prepares for Battle

The European Parliament’s lead negotiators published their draft report in late June. Two committees share responsibility for the legislation: the Committee on Industry, Research and Energy, led by Aura Salla of the European People’s Party, and the Committee on Civil Liberties, Justice and Home Affairs, led by Marina Kaljurand of the Socialists and Democrats. The Legal Affairs Committee and the Internal Market and Consumer Protection Committee also have advisory roles and have produced draft opinions. 

The joint report reflects what Salla and Kaljurand could agree on quickly. It therefore leaves the most contentious questions untouched, including the European Commission’s proposed clarification of “personal data.” 

That restraint will not last. Parliament members have already filed a large number of amendments, and both Salla and Kaljurand have indicated that they will seek changes on issues omitted from the draft, including the definition of personal data. 

The EDPB Concedes—Then Complicates

The EDPB has published new guidelines on anonymization that bear directly on the Commission’s proposal. Most notably, it finally accepted that whether information counts as anonymous—and therefore falls outside the GDPR—depends on the position of the entity processing it. This is the entity-relative view.

That is broadly the clarification the Commission proposes for Article 4, which defines personal data. But the EDPB wrapped that central point in broader, vaguer qualifications than the Commission likely would have adopted. I will return to those complications shortly.  

The ‘AI Omnibus’ Takes Effect

Meanwhile, a less controversial companion measure concerning the EU AI Act entered into force this week. This “AI Omnibus” postponed several compliance deadlines, prohibited AI-generated child sexual-abuse material and nonconsensual intimate imagery, and expanded the legal basis for using sensitive personal data to detect bias in AI systems. 

Where the GDPR Reform Fight Gets Real

The next stage of the fight will turn less on abstract definitions than on whether the final law gives businesses rules they can actually use. The debate now centers on three familiar trouble spots: what counts as personal data, how cookie consent should work, and whether the GDPR will leave room for AI development. 

The Definition Fight Is Mostly Symbolic 

Reports from the final weeks of negotiations under the Cypriot presidency suggest that several national governments oppose the European Commission’s proposed clarification of “personal data” in Article 4 of the GDPR. The proposal also faces resistance in the European Parliament. 

I support the clarification, but its practical importance is easy to overstate. The EDPB has now formally accepted the entity-relative view: whether information counts as personal data depends on the position of the entity processing it. 

Some will argue that any amendment to Article 4 should include qualifications like those in the EDPB’s anonymization guidelines. Others will say that, because the EDPB has already accepted the entity-relative view, no legislative change is needed. 

That argument gives the EDPB too much credit. Its guidelines offer little operational clarity and often seem designed to make implementation difficult unless one simply assumes that all data is always personal. The more consequential reforms are therefore procedural: the Commission’s proposal would let the Commission, rather than the EDPB, adopt legally binding implementing acts that define concepts such as pseudonymization and personal data. 

Those powers should be stronger. As drafted, compliance with a Commission implementing act would count only as “an element” in the legal analysis. It should instead create a robust presumption of compliance. 

Guidance That Actually Guides

What matters most is whether organizations handling EU data can determine which concrete safeguards place information outside the GDPR because it no longer qualifies as personal data. The EDPB’s anonymization guidelines once again show that the board is institutionally ill-suited to provide that kind of practical direction. 

Substance alone is not enough. The guidance must also carry enough legal force to create a strong presumption that those who follow it are acting lawfully. Polish member of the European Parliament Piotr Müller has proposed an amendment moving in that direction—Amendment 398. 

The joint industry and civil-liberties committee draft does not yet revise the relevant provision, Article 41a. Its broader thrust, along with many of the amendments filed in Parliament, instead points toward preserving or even expanding the EDPB’s role. Brussels may yet respond to unworkable guidance by giving its authors more authority. That would be very on-brand. 

Cookies: Pick Your Poison

The weakest part of the Commission’s proposal concerns cookie consent under Article 5 of the ePrivacy Directive. National governments were understandably unimpressed, and the Cypriot presidency reportedly removed the provision from its final draft. 

Ireland has reopened the question, at least in part. It reportedly asked governments whether the law should expand the list of activities that do not require consent, though it is unclear whether the questionnaire also addressed consent managed through browser settings. 

The joint committee draft leaves the Commission’s proposal intact. Other parliamentary amendments cover nearly every imaginable option: replacing the ePrivacy rule with ordinary GDPR standards, adding more exemptions, narrowing the proposed exemptions, deleting browser-level consent, or expanding it. The one thing Parliament appears to agree on is that cookie banners have not yet consumed enough legislative attention. 

AI Rules Back in Play

The Commission’s proposal contains two provisions of particular importance to AI. Article 88c would clarify that legitimate interests may provide a lawful basis for developing and operating AI systems. Article 9(2)(k) would create a limited exception for sensitive personal data that appears incidentally in AI training datasets. 

In my early June comments, I noted that, in the Council drafts, “Article 88c has disappeared from the operative text, but much of its substance survives in Recital 33a,” while “Article 9(2)(k) survived, albeit in narrowed form.” Under the Irish presidency, both issues appear to be back on the table. 

As with cookie consent, the joint committee draft proposes no changes to these provisions. Other amendments run in opposite directions, with some seeking to delete the AI provisions and others seeking to broaden them. 

I am watching Article 9(2)(k) particularly closely because it may be the Commission’s most immediately consequential reform. As I wrote in the International Center for Law & Economics’ (ICLE) March comments: 

Article 9(2)(k) addresses a practical constraint of large-scale AI training: special categories of personal data will inevitably appear in training datasets despite efforts to exclude them. … Perfect ex ante filtering is technically impossible. … Without Article 9(2)(k), controllers face a binary choice: guarantee perfect exclusion of special-category data or abandon AI training in the EU.

Now Comes the Brussels Waiting Game

The European Parliament’s July 15 amendment deadline produced more than 1,000 proposed changes, so the initial joint committee report settles very little. The lead negotiators will spend the autumn bargaining over compromise amendments while awaiting a targeted impact assessment. At this pace, Parliament is unlikely to adopt its negotiating position before February 2027.

The EDPB’s consultation on its anonymization guidelines closes Oct. 30. That process may clarify the board’s position, though recent history counsels against expecting clarity to arrive unaccompanied by qualifications.

The Council may move faster. Ireland now has six months to do what Cyprus could not: secure an agreed position among national governments. Much will depend on how it uses the responses to its questionnaire. The next compromise text should show whether Ireland is serious about simplifying the GDPR—or merely preparing a more elegant route back to the same thicket.

The post Brussels Tries to Fix the GDPR Without Making It Worse appeared first on Truth on the Market.

That two-way relationship will take center stage tomorrow, when the U.S. Senate Commerce Subcommittee on Telecommunications and Media holds a hearing titled “Intelligent Networks: Powering Artificial Intelligence and Transforming Communications.” Networks carry the traffic AI generates, while AI is beginning to reshape how those networks are built, managed, and shared. 

The hearing will likely cover the familiar ground of investment, deployment, and network architecture. Those questions matter. But each rests on a more basic input that rarely gets equal billing: radio spectrum. 

A new International Center for Law & Economics (ICLE) issue brief examines that overlooked foundation in detail. It is the second installment in ICLE’s series on the infrastructure underpinning the next race for AI leadership. 

The Airwaves Won’t Manage Themselves 

Every wireless service depends on reliable access to radio frequencies where competing signals do not cause harmful interference. Spectrum is not scarce in the ordinary sense; nature does not issue a fixed number of frequencies and then close the shop. But physics limits how much data can travel over a given band, and the law limits who may use particular frequencies and under what conditions. 

Those constraints shape the performance of every wireless service. That makes spectrum policy a critical part of infrastructure policy for AI and other next-generation applications. 

Meeting growing data demand does not require tearing down the spectrum-management system and starting over. It does require targeted reforms. Policymakers should preserve a balanced mix of licensed, unlicensed, and shared access to meet the differing needs of AI, augmented reality (AR), and virtual reality (VR). 

They should also move underused federal spectrum into commercial use more quickly by streamlining the cumbersome Spectrum Relocation Fund, improve coordination between the Federal Communications Commission (FCC) and National Telecommunications and Information Administration (NTIA) around a common technical record, and engage more coherently in the international bodies that shape global equipment markets. 

Most promising for a hearing on intelligent networks, policymakers should replace reflexive worst-case interference analysis with risk-informed methods. AI can help here, too. Tools such as spectrum digital twins—virtual models that simulate how networks and signals interact—can estimate interference more realistically and allow more users to share the same frequencies safely. 

New Technologies Need More Than One Lane

AI, AR, and VR will all increase the load on wireless networks, but they will not place the same demands on them. Immersive AR and VR require sustained capacity and extremely low latency—the delay between a user moving her head and seeing the image respond. AI traffic tends to arrive in bursts and, unlike most consumer applications today, can strain upload capacity as well as download capacity. No single frequency band or access model can meet all those needs. 

That makes a layered approach essential. Low-band spectrum travels long distances and penetrates buildings well, but it carries relatively little data. Mid-band spectrum offers a useful balance of coverage and capacity, making it the workhorse for wide-area AR and cloud-rendered VR. The unlicensed 6 GHz band is a natural fit for immersive traffic that remains indoors and travels only short distances. Millimeter-wave spectrum can deliver enormous capacity in dense venues and campuses. Low-Earth-orbit (LEO) satellites add another layer by reaching places that terrestrial networks cannot. 

The rules governing access matter as much as the frequencies themselves. Exclusive licenses give carriers the interference protection and certainty needed to justify billions of dollars in network investment. Unlicensed spectrum provides the permissionless, low-cost capacity that supports Wi-Fi and a connected-device market that generates trillions of dollars in annual value. Dynamic sharing, which coordinates users in real time to prevent harmful interference, opens frequencies that cannot be fully cleared of existing users. 

Each model serves a distinct purpose. Tilt too far toward any one of them, and policymakers will either leave valuable capacity idle or deprive some applications of the access they need. 

The larger problem is that the U.S. system for making spectrum available moves at bureaucratic speed. The FCC governs commercial use, while NTIA manages federal use, and no clear final arbiter exists when they disagree. Much of the prime mid-band spectrum needed for AI and immersive applications remains assigned to federal incumbents, especially the Department of Defense (DOD). 

The One Big Beautiful Bill Act restored the FCC’s auction authority and created an 800 MHz spectrum pipeline. But while Congress can set a target, it cannot force an agency to relinquish frequencies. If the federal portion of that pipeline stalls, pressure will shift to productive commercial, shared, and unlicensed bands—including the 6 GHz band and the Citizens Broadband Radio Service (CBRS), which Congress declined to protect. 

A hearing about powering AI should therefore ask a basic question: Can the reallocation process actually deliver the spectrum these applications will need? 

AI Can Help Spectrum Think Smarter

AI is not just another source of network demand. It can also help networks manage competing radio operations more efficiently—and more intelligently. 

Machine-learning systems can identify unused capacity, predict interference, and adjust channel access in real time. That allows secondary users to operate without disrupting incumbents, meaning users already authorized to occupy the band. Research on cognitive-radio techniques for 6G has found meaningful gains in spectral efficiency—the amount of data transmitted over a given slice of spectrum—and reductions in interference. 

As these methods mature and regulators gain confidence in them, spectrum policy can move beyond the blunt question of whether one user should receive exclusive access to a band. The better question is how efficiently several users can share it. 

The management gains may matter as much as the added capacity. For decades, agencies have relied on deterministic, worst-case interference analysis. These models often assume that every transmitter operates at maximum power, in the least favorable location, under the most adverse conditions, all at once. That caution imposed few costs when spectrum was lightly used. In today’s crowded bands, it can block productive entry based on scenarios that are technically possible but vanishingly unlikely—and give incumbents a potent weapon against competition. 

A risk-informed approach asks better questions: What interference could occur? How likely is it? How serious would the consequences be? 

AI makes that approach more practical. A spectrum digital twin—a continuously updated virtual model of a real radio environment—can run thousands of simulations using realistic combinations of device locations, power levels, and signal conditions. The result is a probabilistic assessment of interference risk, rather than a single alarming hypothetical. 

The FCC already used probabilistic analysis when it opened the 6 GHz band, and the D.C. Circuit upheld that approach as a reasonable and legally defensible basis for spectrum policy. AI-driven modeling could make such analysis the rule rather than the exception. It would preserve robust protections for safety-of-life and national-security systems, where the stakes are genuinely high, while making room for new services elsewhere. 

That is what an intelligent network should look like: not merely a pipe carrying AI traffic, but a system in which automated coordination, real-time sensing, and data-driven interference analysis allow more users to operate safely in the same frequencies. 

Smart Networks Need Smarter Spectrum Rules

If the subcommittee wants networks capable of powering AI, it should focus on five reforms. 

First, regulators should preserve a balanced mix of licensed, unlicensed, and dynamically shared spectrum, and judge allocations by their total economic value—not auction revenue alone. Second, Congress should streamline the Spectrum Relocation Fund so NTIA can study candidate bands earlier and move underused federal spectrum into commercial use faster. 

Third, the FCC and NTIA should continue building a common technical record so disagreements emerge early, rather than erupting at the end of a proceeding. Fourth, regulators should make risk-informed interference analysis the default for spectrum reallocation and sharing. 

Finally, the United States should present coherent positions at the International Telecommunication Union’s (ITU) World Radiocommunication Conference (WRC). That forum shapes the technical standards and global equipment markets on which future networks depend, and China increasingly treats it as an arena for strategic competition. 

Clear direction from the White House will be necessary to align agencies whose missions and incentives often pull in different directions. But the central point is simple, and the July 30 hearing is well positioned to make it: Intelligent networks require intelligent spectrum policy. 

Spectrum is the invisible infrastructure of the AI era. The United States cannot lead at digital speed while governing the airwaves at bureaucratic speed.

The post The AI Boom Still Runs on Old-Fashioned Airwaves appeared first on Truth on the Market.

State attorneys general play a valuable role when harms are local or federal investigators miss key facts. But serial challenges to nationally integrated conduct turn that safeguard into a standing invitation to relitigate.

The result is a system in which no settlement fully settles and no clearance fully clears. That weakens federal enforcement, rewards the most aggressive jurisdiction, and lets local politics reshape national markets.

State antitrust should supply evidence, police local misconduct, and check genuine federal failure. It should not operate as an endless appeals process with 51 possible vetoes.

When Federalism Becomes a Veto Stack

American antitrust federalism has real virtues. State attorneys general can uncover local cartels, represent residents whose individual claims may be too small to pursue, and alert federal officials to facts a Washington-centered investigation might miss. States can also provide a useful check when federal enforcers overlook a genuine competitive harm.

Those benefits do not require subjecting every national transaction or business practice to 51 independent and potentially conflicting antitrust policies.

The growing problem is sequential contradiction. The U.S. Department of Justice (DOJ) or Federal Trade Commission (FTC) investigates a national matter, settles it or closes the investigation, and a coalition of states then tries to keep the same transaction or conduct in court. The firm has not received a federal answer. It has merely cleared the first gate.

This “veto stack” turns federalism from a source of evidence into a mechanism for repeated litigation.

In an earlier Truth on the Market commentary, I argued for a sounder division of labor (see also related work by Babette Boliek). State enforcement should complement well-grounded federal initiatives and focus scarce resources on competitive harms particular to a state.

That account also recognizes that public-choice pressures and rent seeking can distort decisions at both levels of government. The point is institutional, not partisan. The proper allocation of authority is the one that best reduces the combined costs of anticompetitive conduct and mistaken intervention.

Federal Clearance, State-Sized Asterisk

Recent litigation reveals several forms of state-federal divergence. They should not be forced into a single narrative because each raises a different institutional concern. Together, though, they suggest that federal clearance increasingly comes with an asterisk.

Live Nation offers the strongest warning against claiming that continued state litigation is always wasteful. The DOJ announced a proposed settlement shortly after trial began, but most state plaintiffs stayed in the case and won a jury verdict.

One could still view that outcome as a policy error. Live Nation argues that combining promotion, venue operations, and Ticketmaster’s ticketing services creates an efficient vertical integration that lowers costs. But even assuming the jury reached the right result, one verdict cannot resolve the broader institutional question.

Firms bear the ex ante costs of overlapping authority in every case, including those in which a second prosecution produces no offsetting consumer benefit. Conflicting remedies can also persist long after a court determines liability.

HPE/Juniper raises a different concern. The DOJ first challenged the $14 billion acquisition, then settled for the divestiture of HPE’s Instant On business and licensing commitments covering Juniper’s artificial-intelligence (AI) operations source code.

The states did not file a separate Clayton Act suit. Instead, 12 states and the District of Columbia intervened in the Tunney Act proceeding, which allows courts to review federal antitrust settlements, and urged the court to reject or closely examine the decree.

John Yun’s analysis focused on a key aspect of competition to which the original complaint gave too little weight. Combining Juniper’s AI-powered networking capabilities with HPE could create a stronger rival to Cisco and increase competition. Greater international competitiveness could also advance national security interests.

The state intervention extended uncertainty over a completed integration whose competitive rationale depended on AI networking, scale, and the prospect of a more credible third platform.

Nexstar/Tegna and Paramount Skydance/WBD go further. In Nexstar, state plaintiffs sued after federal regulators approved the deal and the transaction closed.

In Paramount, the DOJ publicly ended an eight-month investigation after concluding that the asserted theories appeared to confuse harm to competitors with harm to consumers. A 12-state coalition filed a federal antitrust suit the next month.

The merits of those challenges remain distinct. The institutional message is harder to miss. A federal resolution may merely open the next round.

The High Cost of One More Veto

Antitrust enforcement is an exercise in managing error costs. False negatives allow market power to persist. False positives deter conduct that lowers costs, improves products, or creates new competitive constraints. The goal should be to minimize the expected social costs of both errors, along with the administrative and compliance costs of enforcement—not to maximize cases, remedies, or the number of officials with a vote.

Sequential state prosecution increases those costs in at least four ways.

First, it erodes the settlement value of federal enforcement. A firm cannot reliably exchange concessions for closure when non-settling states remain free to pursue conflicting relief. That uncertainty raises the price of settlement, weakens the government’s ability to secure efficient remedies, and encourages firms to litigate defensively.

Second, it delays integration. Merger efficiencies often depend on timing. Engineers leave, product road maps diverge, capital budgets freeze, and customers postpone adoption while a transaction’s legal status remains uncertain.

Third, a veto stack magnifies uncertainty. The relevant business question becomes whether any politically prominent jurisdiction might attack the conduct under a more elastic standard, rather than whether the conduct satisfies a coherent consumer-welfare rule. The most aggressive enforcer then determines expected liability, regardless of which agency has the best information.

National companies respond predictably. They design their conduct around the strictest state rule, abandon marginal investments, or avoid business models whose legality requires a fact-intensive assessment of efficiencies.

Fourth, fragmented enforcement creates remedy externalities. A state may seek divestiture, access mandates, data sharing, compulsory licensing, or conduct restrictions whose costs fall largely beyond its borders. State officials receive political credit for a visible intervention, while consumers, workers, and shareholders nationwide absorb the lost scale or innovation. The jurisdiction making the decision does not bear the full national cost.

These risks grow sharper in high technology. Dynamic competition depends on uncertain investments, complementary assets, learning effects, and rapid quality improvements that static market shares often measure poorly. My dynamic-consumer-welfare approach therefore focuses on innovation and long-run consumer gains.

The International Center for Law & Economics’ (ICLE) work on state laws with national spillovers makes a related point. State experimentation becomes costly when one jurisdiction exports the consequences of its policies to the interstate economy. Requiring every state to grow comfortable with an unfamiliar technology before firms may proceed will predictably protect incumbents and punish experimentation.

When ‘AG’ Means ‘Aspiring Governor’

Why do national technology firms make such tempting targets? Public-choice analysis offers a simple answer without accusing anyone of corruption or bad faith.

State attorneys general are elected officials or politically prominent generalists. Large technology cases generate headlines, coalition leadership, fundraising material, and a platform for higher office. Competitors, organized interest groups, and plaintiffs’ lawyers also have strong incentives to provide complaints, studies, and political support.

The beneficiaries of innovation are harder to organize. Future customers, small businesses, and users who value free or improved services rarely form a ready-made constituency.

Empirical research cannot prove that any particular case rests on political motives. It can, however, identify the incentives built into the system. Colin Provost examined when “AG” becomes shorthand for “aspiring governor” and found a relationship between multistate litigation and later campaigns for higher office. John Dove’s work on state antitrust enforcement likewise considers electoral cycles and institutional political pressures.

Those findings counsel restraint. Broader discretionary standards do not produce neutral public-interest enforcement by magic. They increase the rewards for political entrepreneurship.

Technology markets sharpen the problem. Their products are complicated, their alleged harms translate easily into populist rhetoric, and their benefits often appear as better quality, lower latency, stronger security, faster model development, or future competition rather than lower prices today.

A prosecutor can point to size, data, vertical integration, or an unhappy rival. Showing that integration improved performance or created a stronger competitor takes more work. The political reward arrives when the complaint is filed. The innovation lost to delay rarely makes the evening news.

Fifty-One Rulebooks for One National Market

Conflicting enforcement becomes more dangerous when states also apply different substantive rules.

For roughly four decades, the consumer-welfare framework has disciplined antitrust by asking whether challenged conduct is likely to reduce output, raise quality-adjusted prices, suppress innovation, or otherwise harm competition. The framework does not grant firms laissez-faire immunity. It separates competition law from a broader power to redistribute economic advantage among firms and political constituencies.

Several state initiatives point toward wider and less predictable mandates. New York’s Twenty-First Century Antitrust Act, which passed the state Senate in May and remains pending in the Assembly, would create an abuse-of-dominance regime, restrict certain vertical conduct, expand premerger requirements, and authorize private lawsuits. My critique of an earlier version warned that such a standard could protect competitors at consumers’ expense and interfere with national policy.

California’s pending AB 1776 would expand single-firm liability under the Cartwright Act. Babette Boliek’s recent ICLE critique argues that the bill would replace objective consumer-welfare analysis with protection for favored market participants. New Jersey’s S451 would make certain uses of algorithmic rent-coordination systems violations of the New Jersey Antitrust Act.

These proposals differ in scope and may change or fail. Together, they show growing support for state-specific theories that could govern conduct far beyond state borders.

The expertise concern is comparative, not absolute. Many state antitrust lawyers are excellent, and federal agencies make mistakes. But the DOJ and FTC employ larger teams of economists, technologists, merger specialists, and lawyers focused on national markets. They also retain more institutional knowledge from repeated investigations.

A state coalition formed for a prominent case may rely on federal evidence and outside counsel while pursuing a different political objective. When a dispute concerns AI infrastructure, cloud architecture, enterprise networking, algorithmic pricing, or multisided platforms, analytical mistakes become especially costly.

Broad fairness or dominance standards also invite rent seeking. Rivals can repackage requests for protection as antitrust complaints. Regulated intermediaries can seek rules that preserve their position. Local interests can impose costs on consumers nationwide.

Recent Truth on the Market commentary has described this tendency as the “nanny state” substituting regulators’ preferences for demonstrated consumer harm. The law & economics objection does not dismiss every nonprice concern. It simply observes that a standard detached from measurable competitive effects gives too much discretion to officials who face uneven political incentives.

Congress Holds the Preemption Pen

Reform cannot assume that Washington merely lends states their antitrust authority. States may sue under their own statutes, and Congress has expressly authorized state attorneys general to bring parens patriae actions for damages under federal antitrust law. These actions allow a state to seek recovery on behalf of its residents.

In California v. ARC America, the Supreme Court held in 1989 that federal limits on indirect-purchaser recovery did not preempt state laws permitting such claims. The Court reasoned that federal antitrust law did not occupy the entire field and that federal policy alone did not dictate what remedies states could authorize.

ARC America therefore poses a serious obstacle to casual claims of implied preemption. But the decision does not require Congress to preserve overlapping state remedies forever. It interpreted the statutes Congress had enacted.

Congress may regulate interstate commerce and expressly preempt conflicting state rules, so long as it speaks clearly and respects constitutional limits against commandeering state governments. The cleaner approach would establish a federal rule for private conduct and transactions in national markets, backed by an express preemption clause and a carefully drawn exception for genuinely local enforcement. Congress would govern private conduct rather than tell states how to govern.

Courts have less room to act without Congress. Modern dormant Commerce Clause doctrine generally resists invalidating nondiscriminatory state laws merely because they affect conduct beyond the state. The Supreme Court reinforced that caution in National Pork Producers Council v. Ross.

Conflict preemption remains possible when compliance with both state and federal law is impossible or when state law obstructs a clearly stated federal objective. Yet ARC America makes that objective difficult to infer from antitrust policy alone. Judicial reform would therefore require a clearer signal from Congress or a modest doctrinal shift toward giving nationally uniform federal resolutions greater finality.

How to Defuse the Veto Stack

The answer is not to push states out of antitrust enforcement. States can uncover local harms, contribute evidence, and challenge a plainly deficient federal process. But nationally integrated conduct needs a federal resolution that carries more weight than an opening bid.

A workable reform agenda should preserve state participation while limiting contradictory remedies and serial litigation. Congress can clarify when federal decisions preempt state action, courts can give greater weight to reasoned federal assessments, and federal agencies can improve coordination and transparency. AI markets make those reforms especially urgent because even a nominally local injunction can alter a national product.

Make Federal Resolution Mean Actual Resolution

Congress should make federal antitrust resolutions final for conduct or transactions with substantial, nonincidental effects across state lines. Once the DOJ or FTC completes a defined review and issues a reasoned disposition, federal substantive standards should govern exclusively.

A qualifying consent decree, litigated judgment, or published closing statement could preempt later state actions seeking inconsistent prospective relief. States should remain free to participate in the federal investigation, submit evidence, seek compensation for residents, and challenge conduct whose competitive effects are predominantly intrastate.

The dividing line should turn on function, not size alone. Relevant factors could include nationwide pricing or product design, a multistate network, integrated data or infrastructure, or a remedy that would necessarily alter operations nationwide.

A savings clause should preserve state cases involving local bid rigging, regional cartels, occupational restraints, hospital or funeral-home conspiracies confined to one state, and small mergers whose assets and customers are overwhelmingly local. The governing principle should be subsidiarity, which assigns authority to the lowest level of government capable of accounting for the relevant costs and applying adequate expertise.

Preclusion should apply only after a transparent federal process. The DOJ or FTC should explain the theories investigated, the competitive evidence considered, and why the disposition protects consumers.

States could receive a brief consultation period and limited judicial review of whether the statutory conditions were satisfied. They should not receive a second trial on the merits. Congress could also require periodic reports on matters in which a federal resolution displaced a proposed state action.

Police Conflicting Remedies

Even without full field preemption, Congress could direct courts to displace state remedies that materially frustrate a federal resolution of national conduct.

A state damages claim based on a distinct local injury may coexist with a federal decree. A state injunction requiring divestiture after the DOJ accepted a licensing remedy presents a different problem. The question should be whether the state remedy conflicts with the federal resolution, not whether state officials merely disagree with it.

Courts should ask whether the requested relief would undo the settlement bargain, destroy efficiencies the federal disposition preserved, or force changes to a nationwide operating plan that federal enforcers approved.

Courts should remain cautious without becoming passive when Congress has not spoken. If a state brings a federal antitrust claim after the DOJ or FTC issues a detailed competitive assessment, courts should give that assessment substantial weight on market definition, efficiencies, and remedy design.

That approach would preserve state standing while reducing the anomaly of one sovereign obtaining federal relief that defeats an expert federal agency’s stated national policy.

Put the Federal Case on the Record

The most immediate reform requires no new statute. The DOJ already files statements of interest in private and state litigation. Both the DOJ and FTC should adopt a published protocol for intervening when a state case threatens a federal settlement, relies on a theory the federal agency rejected, or seeks a remedy with substantial interstate spillovers.

The filing should set out the federal consumer-welfare analysis, explain the national costs of conflicting relief, and address technical questions that a generalist court might otherwise hear only through dueling experts.

Such filings would be especially useful in Tunney Act proceedings. States should have an opportunity to be heard, but the public-interest review should not become a chance to retry the government’s entire case or replace a federal settlement with a coalition’s preferred remedy.

A clear federal account of uncertainty, litigation risk, innovation benefits, and remedial tradeoffs would help courts distinguish genuine defects from political disagreement.

Make States Explain the Encore

Federal officials should say plainly that cooperation is welcome, but serial prosecution of nationally integrated conduct carries real economic costs.

Congressional hearings could require federal and state enforcers to explain their competing theories, staffing, outside-counsel arrangements, and estimates of interstate spillovers. Greater transparency would force state officials to defend both the alleged harm and the added value of a separate action after federal review.

The DOJ and FTC could also condition joint investigations on ex ante coordination agreements governing information sharing, settlement consultation, and remedy design. A state would remain sovereign and free to withdraw. It should not be able to use the federal investigation, then surprise the parties with a conflicting national remedy.

Published best practices could encourage state offices to conduct cost-benefit analysis and reserve independent action for demonstrable local harm or a clearly deficient federal process.

Give AI One National Rulebook

AI-related conduct will test this system first. Training, cloud computing, chips, networking, data centers, model distribution, and enterprise deployment cross state lines by design. A state injunction governing model access, data use, interoperability, or vertical integration could reshape a national product even when issued under a nominally local statute.

Congress could establish expedited federal review and presumptive federal primacy for AI cases with substantial interstate effects. States would retain their traditional authority over fraud, privacy, contracts, and local cartels.

Federal primacy should preserve vigorous enforcement against exclusion, collusion, and acquisitions likely to harm consumers. The error-cost analysis must also account for dynamic global competition.

As Mario Zúñiga recently argued, antitrust accounts that omit China ignore a major rival in AI development. HPE/Juniper likewise shows how integrating AI-enabled networking can create a stronger challenger. The United States will struggle to out-innovate strategic competitors if every national technology investment depends on satisfying the most economically untethered state theory.

Federal Primacy, With Guardrails

The strongest objection is that federal enforcers can be captured, politicized, or simply wrong. Live Nation’s jury verdict gives that concern real force. A system that turned any cursory federal closing letter into blanket immunity would merely trade one institutional failure for another.

Federal primacy should therefore depend on a credible process. Federal enforcers should conduct an adequate investigation, provide a reasoned public explanation, disclose the logic behind any remedy, and submit consent decrees to judicial review where required.

Preemption also should not erase state damages claims on behalf of residents where federal law expressly preserves them, unless Congress clearly chooses otherwise. Prospective injunctions and efforts to block transactions create the largest national spillovers and present the strongest case for uniform federal treatment.

States should retain broad authority to compensate localized injuries, prosecute local conspiracies, and participate in federal cases. The governing principle is matching jurisdiction to the geographic scope of the market and the remedy, rather than centralizing authority for its own sake.

A more disciplined federalism could also improve state enforcement. State offices would have more resources for procurement cartels, local licensing restraints, unlawful noncompete agreements, health-care consolidation with genuinely regional effects, and anticompetitive state regulation. In those matters, local knowledge offers a genuine advantage rather than a vehicle for imposing one state’s policy nationwide.

Federalism Without the Endless Encore

Antitrust federalism should improve information and accountability. It should not ensure that a national firm can never obtain repose.

When states continue litigating after a reasoned federal settlement or clearance, the costs extend far beyond legal fees. Integration stalls, investment falls, remedies conflict, settlements lose value, and unfamiliar innovations face a built-in handicap. Those costs hit hardest in AI and other technologies where speed, scale, and experimentation shape long-run consumer welfare and national competitiveness.

The answer is neither abolishing state antitrust nor treating federal expertise as infallible. Congress should preserve state authority over genuinely local restraints while restoring federal primacy over conduct with substantial interstate effects. Courts should police conflicting remedies when federal policy is clear. The DOJ and FTC should coordinate earlier, file statements of interest, and explain when state actions threaten consumer welfare. Congressional oversight should expose the incentives behind serial prosecution.

A national market cannot function efficiently with 51 final decision makers. States should lead where the harm and expertise are local. Federal enforcers should answer for markets and remedies that are national.

Federalism should divide the work—not multiply the vetoes.

The post Too Many Cooks Spoil the Settlement appeared first on Truth on the Market.

For four decades, the FCC has barred any company from owning television stations that collectively reach more than 39% of U.S. television households. The rule reflects an older theory of broadcast regulation—one that treats the airwaves as a scarce public resource and promotes competition, localism, and viewpoint diversity through bright-line ownership limits rather than case-by-case review.

That theory once had an intuitive logic. When the modern cap emerged, most Americans got their news and entertainment from stations affiliated with the “Big Three” networks. Limiting any owner’s national footprint could plausibly prevent too much editorial influence from accumulating in too few hands.

That media world has vanished. Broadcasters now compete not only with one another, but also with streaming services, virtual multichannel video programming distributors (vMVPDs), podcasts, and social platforms whose national and global reach dwarfs anything a station group could assemble under the 39% cap. The rule now binds the competitors least able to bear it while leaving their largest rivals untouched.

The FCC has signaled that it intends to repeal the cap and review broadcast consolidation case by case. As a matter of competition policy, that move is overdue. As a matter of law, it is messier. Congress wrote the 39% figure into an appropriations statute, raising the question of whether the FCC may erase it on its own.

The policy debate also cannot stop at ownership. Broadcast regulation operates as an interconnected system. Its other parts include retransmission consent, which governs the terms and fees under which distributors carry broadcast signals; must-carry rules, which can require carriage of qualifying local stations; and FCC standards requiring the parties to negotiate in good faith. Together, these rules divide bargaining power among many of the same companies.

Removing the cap would give larger station groups more leverage over the cable, satellite, and streaming distributors that carry their signals. The retransmission regime was not designed to offset that added power. Repealing the cap while leaving the carriage rules untouched would therefore do less to eliminate a distortion than to move it elsewhere.

The better course is comprehensive reform. Because ownership limits, retransmission consent, and bargaining standards all fall within the FCC’s jurisdiction, the agency should consider them together. Otherwise, repeal may simply reshuffle bargaining power among industry players while consumers keep paying the bill. 

How 39% Became Broadcast’s Magic Number

The national television ownership rule measures concentration by audience reach, not station count. A broadcaster may not hold an attributable interest—a stake large enough to count as ownership under FCC rules—in stations that collectively reach more than 39% of U.S. television households.

The rule grew out of broadcasting’s traditional “scarcity rationale.” Because the electromagnetic spectrum could accommodate only a limited number of signals, regulators treated licensees as stewards of a public resource with a duty to serve the public interest. The FCC therefore pursued competition, localism, and viewpoint diversity through fixed ownership limits rather than reviewing each transaction on its particular facts.

The theory was straightforward. Preventing any company from building a nationwide footprint would preserve a decentralized system of locally accountable stations and keep editorial influence from accumulating in too few hands. That concern carried more weight in the 1980s, when most Americans relied on stations affiliated with the “Big Three” networks for news and entertainment.

The 39% cap is only the latest entry in a much longer regulatory ledger. The FCC first imposed national ownership limits in the 1940s, initially by restricting how many stations one company could own. It later shifted to the audience-reach measure used today.

With the Telecommunications Act of 1996, Congress directed the FCC to raise the cap to 35%. When the agency tried to increase it in 2003 to 45%, Congress again stepped in. The Consolidated Appropriations Act of 2004 changed the figure to 39%, where it has remained.

That history now sits at the center of the legal dispute over whether the FCC may repeal the cap without further action from Congress. Whatever the answer, the number rests on the same premise that has long animated the rule: Limiting a broadcaster’s national reach will protect a diverse, locally oriented media system.

A Broadcast Cap in a Streaming World

In today’s media economy, that premise points in the wrong direction. Local broadcasters compete for audiences and advertising dollars with streaming services, vMVPDs, podcasts, and social platforms whose national and global scale dwarfs anything a station group could assemble under a 39% ceiling. A rule intended to prevent broadcasters from becoming too powerful now risks keeping them too small to compete effectively.

The FCC has often resisted this conclusion by treating another service as a competitor only if it substitutes for broadcasting in every part of the business. Cable and streaming services, for example, may be treated as complements rather than competitors because they do not bid for network affiliations or retransmission-consent fees.

That reasoning is circular. Cable operators do not bid for retransmission-consent fees because federal law puts them on the paying side of that transaction. The regulatory system creates the difference, then cites the same difference as proof that broadcasters and distributors occupy separate markets and deserve different treatment.

Consumer behavior tells a different story. By Nielsen’s measure, streaming now accounts for roughly 48.6% of television viewing time, compared with 21.5% for broadcast. Over-the-air radio’s share of audio listening likewise fell from 46% to 34% between 2018 and 2025 as streaming and podcasts gained ground. A market defined narrowly as “local broadcast television” no longer reflects how audiences choose content or advertisers spend money.

Against that backdrop, lifting the ownership cap would be pro-competitive. Broadcasters can promote localism and viewpoint diversity only if they remain strong enough to offer a meaningful alternative to digital platforms. YouTube and Netflix may reach the entire country. Broadcasters remain frozen at 39% of television households.

Removing that asymmetry would let station groups build some of the scale their digital rivals already enjoy. Scale can make broadcasters more effective competitors, not less. Larger groups can spread the fixed costs of investigative teams, weather forecasting, and production infrastructure across more stations, reducing the average cost of producing local journalism that national platforms cannot easily replicate.

Greater scale could also strengthen broadcasters when they acquire programming and negotiate carriage on the streaming services and vMVPDs that increasingly connect them with viewers. It could help finance the “must-have” local content that distinguishes stations in advertising markets dominated by technology companies.

On those terms, consolidation may preserve localism and viewpoint diversity rather than threaten them. The greatest danger to local broadcasting may not be excessive concentration in midsize markets. It may be the financial collapse of stations that outdated rules prevent from adapting. A cap that keeps broadcasters too small to compete efficiently starves them of investment and cedes more ground to the digital giants the rules leave untouched.

The cap also exposes a deeper problem: the public-interest standard through which the FCC regulates broadcasters. The Communications Act directs the agency to act in the “public interest, convenience, and necessity,” but gives those words no concrete definition.

As Eric Fruits has observed in revisiting Stephen Breyer’s classic critique, such vague standards are a familiar regulatory pathology. They give agencies sweeping discretion while preserving the appearance of rule-bound decision-making. Because no objective test can measure the “public interest,” different commissions can use the same words to justify sharply different results. 

The 8th U.S. Circuit Court of Appeals’ decision in Zimmer Radio shows how that vagueness feeds back into the ownership rules. The court interpreted the statutory term “competition” through the public-interest standard, treating it as a broad grant of authority with few meaningful limits. It then deferred to an FCC market definition that excluded broadcasters’ digital rivals. 

The FCC’s use of the public-interest standard to promote localism and viewpoint diversity may also threaten broadcasters’ First Amendment rights. A standard with few limiting principles allows the agency to regulate content, condition license grants, and demand concessions from merging parties without tying those demands to any harm caused by the transaction. 

Broadcasting’s reduced First Amendment protection rests largely on the spectrum-scarcity rationale of Red Lion, which upheld content obligations that would almost certainly fail if imposed on a newspaper. That rationale has crumbled in an era of digital abundance. The result is an increasingly hard-to-defend system in which the same speech receives full constitutional protection in print, on cable, and online, yet remains subject to special FCC oversight when transmitted by a broadcast station. 

Repealing the cap would not cure every defect in the public-interest standard. But it would, at least, allow the FCC to assess consolidation across the broader media market rather than inside an increasingly artificial broadcast silo. 

Can the FCC Kill Congress’ Cap?

The policy case for repeal, however strong, faces a threshold objection: Congress mandated the 39% figure. Congress placed no limit on the raw number of television stations a company may own nationwide, but it imposed a firm ceiling on their combined audience reach. Whatever the cap’s merits, an agency ordinarily must apply the law Congress enacted until Congress changes it.

Two features of the Consolidated Appropriations Act of 2004 bolster that reading. First, Congress excluded the national cap from the quadrennial-review process, under which the FCC reexamines its other ownership rules every four years. That omission suggests Congress reserved this particular number for itself rather than leaving it to periodic agency reconsideration.

Second, Congress expressly barred the FCC from using its Section 10 forbearance authority to avoid applying the cap. Forbearance allows the agency, in certain circumstances, to stop enforcing regulatory requirements it finds unnecessary. Section 10 applies by its terms to telecommunications carriers and services under Title II of the Communications Act, while broadcast licensing falls under Title III. The cross-reference may therefore reflect imprecise drafting or an abundance of legislative caution. Either way, opponents of repeal argue that Congress’ message was plain: The FCC could not simply set the cap aside.

Recent Supreme Court doctrine makes the agency’s position more precarious. In Loper Bright Enterprises v. Raimondo, the Court retired Chevron deference, under which judges often accepted a reasonable agency interpretation of an ambiguous statute. Courts must now decide for themselves what the statute means, and a reviewing court may reject the FCC’s interpretation outright. 

The major-questions doctrine may present another obstacle. Under that doctrine, courts expect clear authorization before an agency makes a decision of vast economic or political significance. Opponents of repeal could argue that eliminating a nationwide ownership limit qualifies and that Congress never clearly gave the FCC that power. On this view, the cure for an obsolete cap is legislation, not an agency decision to stop enforcing it. 

The FCC nevertheless has a serious argument that Congress never stripped it of authority to revisit the cap. The agency created the national ownership limit under the Communications Act’s broad grant of regulatory authority and revised it repeatedly before Congress intervened. 

When Congress set a 35% cap in the Telecommunications Act of 1996, it directed the FCC to “modify its rules” accordingly. In 2002, the D.C. Circuit held that this language preserved the agency’s discretion to alter or eliminate the cap. The court reasoned that Congress could have written the number directly into the statute had it wished to freeze the limit in place. 

Two years later, presumably aware of that ruling, Congress raised the figure from 35% to 39% while retaining the same “modify its rules” language. It also declined to enact competing House and Senate bills that would have expressly codified the cap. Preserving language that a federal court had already interpreted as leaving the FCC room to act looks less like a withdrawal of authority than a ratification of it. 

The other statutory provisions do not necessarily foreclose that reading. Excluding the cap from mandatory quadrennial review does not automatically eliminate the FCC’s independent authority to reconsider it. Nor does a prohibition on forbearance necessarily prevent the agency from repealing the rule through notice-and-comment rulemaking—the standard process in which an agency proposes a rule, receives public comments, and explains its final decision. Declining to enforce a rule and lawfully rescinding it are not the same thing. 

The demise of Chevron may even cut in the FCC’s favor. Loper Bright requires courts to determine the statute’s best reading, but it also recognizes that Congress may delegate policy choices to agencies. A reviewing court could conclude that Congress directed the FCC to modify its ownership rules while leaving the agency discretion over what those rules should become. 

The FCC would receive no automatic deference on the statute’s meaning. But a court could still find that Congress gave the agency authority to make the choice. The end of deference is not the end of delegation.

Pull One Lever, Move the Whole Machine

Lifting the cap cannot be treated as a stand-alone reform. The ownership rules, Title VI’s retransmission-consent and must-carry regime, and the FCC’s good-faith bargaining standards all govern the same firms and shape leverage across the same market. Changing one part while leaving the others untouched does not eliminate a distortion so much as relocate it.

The link between ownership and retransmission consent is bargaining power. Retransmission fees have grown to rival advertising as a source of broadcaster revenue, and a station group’s leverage rises with the size of its footprint. Removing the cap would therefore give consolidated broadcasters more leverage over the cable, satellite, and vMVPD distributors that carry their signals.

The retransmission-consent regime was not designed with that leverage in mind. When Congress created it, cable operators were widely viewed as local monopolists. Today, larger station groups can use their scale to demand higher carriage fees. Distributors may respond by passing those costs to subscribers or dropping smaller, independent networks to contain expenses.

The regime’s original rationale has also weakened on its own terms. Congress adopted mandatory carriage in the 1992 Cable Act to address the “cable bottleneck”—the fear that monopoly cable systems would refuse to carry local stations. But as the D.C. Circuit recognized in Comcast Corp. v. FCC in 2009, cable operators no longer possess the bottleneck power that animated Congress in 1992. Broadcasters can now reach viewers through their own websites, apps, and streaming services.

Once the bottleneck disappears, the case for compulsory carriage—and for the elaborate machinery of fee negotiations and blackouts layered on top of it—becomes much weaker. Repealing the ownership cap would not create that mismatch, but it would magnify it by removing one constraint on broadcasters while preserving a carriage regime built for another era. 

Revisiting retransmission consent could take several forms. The cleanest option would be to phase out both retransmission consent and must-carry, treat broadcasters like other content owners, and let copyright law and voluntary contracts govern distribution. As Geoffrey Manne argued in congressional testimony in 2013, that approach would also end the regulatory asymmetry between broadcasters and their streaming competitors. 

Short of repeal, the FCC could tighten its good-faith bargaining rules to prevent parties from using scale merely to delay negotiations, posture, or bundle unrelated demands. It could also limit automatic fee-escalation clauses that allow each acquisition to ratchet up rates across an entire portfolio. 

The agency might also consider final-offer, or “baseball-style,” arbitration for disputes during high-value programming windows. Each side would submit its best offer, and an arbitrator would select one rather than splitting the difference. Because an unreasonable proposal risks losing outright, the process discourages extreme demands and reduces the chance that viewers become collateral damage in blackout brinkmanship.

The broader point is simple: Broadcast regulation will shift bargaining power somewhere. The relevant question is not which industry gains leverage, but whether consumers receive lower prices, better programming, or more reliable service.

Repealing the cap by itself may simply transfer revenue from distributors to larger broadcasters without producing any consumer benefit. Because ownership rules, carriage rules, and bargaining standards all fall within the FCC’s jurisdiction and govern the same products and firms, this is one of the rare cases in which comprehensive reform is not just desirable, but achievable.

Don’t Stop at 39%

The national ownership cap was built for a media market that no longer exists, and the FCC is right to seek its retirement. But the case for repeal is also a warning against half-measures.

Whatever a court ultimately decides about the Commission’s authority, the broader lesson is clear: Broadcast regulation is a system, not a stack of unrelated rules. Ownership limits, retransmission consent, and bargaining standards all shape leverage among the same firms. Pull one lever, and the others move.

If the FCC repeals the cap but leaves the carriage rules untouched, it will not have deregulated so much as redistributed bargaining power. Larger broadcasters will gain leverage that the retransmission regime was never designed to check, with no guarantee of lower prices, better programming, or more reliable service for consumers.

The Commission should not merely lift the cap. It should finish the job.

The post Pulling the 39% Thread: Why the FCC Must Fix More Than the Broadcast Cap appeared first on Truth on the Market.

The International Center for Law & Economics (ICLE) submitted comments in response to the FTC’s request for input, and this post is, in part, a digest of those comments. The statement gets some important things right, including the risks of excessive regulation and a patchwork of state laws. But it offers little concrete guidance on deception, leans on dubious assumptions about consumer expectations, wanders into constitutionally protected editorial judgments, and treats federal preemption as more wish than doctrine. 

In other words, the AI Policy Statement needs considerable work. Then again, perhaps guiding enforcement was never quite the point. 

But first, some context. 

Just Asking Questions—and Testing the Limits

Way back in February 2025, the FTC issued a Request for Public Comment Regarding Technology Platform Censorship (RFI). There, as you may recall, the agency explained: 

FTC staff is interested in understanding how consumers have been harmed—including by potentially unfair or deceptive acts or practices, or potentially unfair methods of competition—by technology platforms that limit users’ ability to share their ideas or affiliations freely and openly. 

FTC staff was, as the kids say, “just asking.” Then again, perhaps it was not just asking. The RFI sought comment on whether technology platforms’ content-moderation practices might violate either the consumer-protection prong of Section 5 of the FTC Act—unfair or deceptive acts or practices (UDAP)—or its competition prong—unfair methods of competition (UMC). Some questions seemed loaded—and loaded for bear, at that. Comments from FTC leadership describing tech “censorship” as “un-American” did little to dispel that impression. 

ICLE’s comments on the RFI raised those concerns, as did submissions from the Program on Law, Economics, and Privacy at George Mason University’s Antonin Scalia Law School, the Foundation for Individual Rights and Expression (FIRE), the American Enterprise Institute (AEI), and the Cato Institute. ICLE offered “reasons why both the law and underlying economics support limiting how the FTC acts in response to the information gained in this RFI,” while acknowledging that “a call for diverse public comments is likely the best means at the agency’s disposal to gather preliminary information, and that no subsequent formal economic study or law-enforcement action is implied by such an inquiry.” 

Still, our comments were heavy on the limits and brief on the allowance. Some addressed the potential use of the FTC’s UDAP authority against technology platforms’ content-moderation policies and practices. The rest considered whether the FTC could use its UMC authority—or the federal antitrust laws more broadly—to challenge content moderation. 

That emphasis reflected a multipronged initiative by both the FTC and the U.S. Department of Justice’s (DOJ) Antitrust Division suggesting that content moderation might violate the antitrust laws. Beyond the RFI, FTC Chair Andrew Ferguson made informal comments about “tech platform censorship,” while Dina Kallay, deputy assistant attorney general for antitrust at DOJ, offered related remarks. 

DOJ also filed a statement of interest in a private antitrust suit brought against several major news publishers by Children’s Health Defense and other plaintiffs. U.S. Health and Human Services Secretary Robert F. Kennedy Jr. previously chaired Children’s Health Defense. The suit itself seemed dubious, and The Wall Street Journal was no more impressed by DOJ’s statement of interest, which it called “bizarre.” 

To be clear, our position was not that content-moderation policies could never support a viable antitrust or UDAP claim. It was that the space for such claims was narrow. Among the hurdles, platforms’ content-moderation decisions are themselves protected by the First Amendment to the U.S. Constitution. 

The Supreme Court’s decision in Moody v. NetChoice seemed especially relevant. The case concerned Florida and Texas laws that sought to regulate “large social media companies and other internet platforms.” As the Court explained, those laws would “curtail the platforms’ capacity to engage in content moderation—to filter, prioritize, and label the varied third-party messages, videos, and other content their users wish to post.” 

That holding did not foreclose every possible challenge to content-moderation policies or practices, but it narrowed the field considerably. Platforms, like newspapers, may still face liability for price fixing or unlawful monopolization in their business dealings, even when their content-moderation decisions qualify as protected speech. 

For more on this Scylla and Charybdis—and the dangerous route between them—we recommend Content Moderation Antitrust, a collection published by Concurrences and edited by Bilal Sayyed, former director of the FTC’s Office of Policy Planning. It includes our article, “Is There an Empty Set at the Intersection of Antitrust and Content Moderation?” (available free here). 

In the year that followed, the issue seemed to fade. Or, at least, neither the FTC nor DOJ took it to court. 

Credit Where It’s Due—Before the ‘But’

Concerns about artificial intelligence (AI)—and proposals to regulate it—are myriad and, in many cases, “out there.” That does not mean there are no legitimate concerns, or that policymakers could not design well-tailored rules to address durable market failures that harm consumers. We simply have not seen many such rules proposed or adopted.

As ICLE’s comments on the AI Policy Statement observe, “[t]he FTC gets much right in this proposed AI Policy Statement.” There is a “but” coming, but credit where it is due.

First, the statement understandably responds to White House executive orders and policy directives. As the Supreme Court recently reminded us in Trump v. Slaughter, the FTC is an executive agency, and the Constitution vests “the executive power” in the president of the United States.

Questions about the proper scope of that power aside, it is entirely appropriate for the FTC and its commissioners to respond to executive orders issued by the White House—and not merely because the Court has held that the president may remove a commissioner at will and without cause.

It is fitting, then, that the AI Policy Statement responds to several White House documents, including an AI Action Plan, a report on AI from the Council of Economic Advisers, and two executive orders (here and here). 

Four Things the FTC Gets Right

Several policy observations in the FTC’s statement seem salutary.

First, we agree that consumers and competition at every level of the AI stack may be better protected by uniform federal policy than by a motley collection of state regulations. That is not because federal law should generally displace state regulation—there is no good reason to think so—or because, to borrow Justice Louis Brandeis’ phrase, a state cannot serve “as a laboratory” for testing novel policies, assuming it collects the right data and conducts careful analysis.

Leave aside, for now, that some state proposals strike us as bonkers. The central problem is that several features of AI weigh against state-by-state experimentation: the rapid pace of AI development, the interstate reach of leading firms and their products, and the nationwide distribution of consumers. State AI policies are therefore likely to produce substantial spillover effects beyond their borders.

A thicket of potentially conflicting state regulations would also add considerable complexity—and other costs—to AI development. We are therefore inclined to agree that the country needs a “national AI framework” that will “protect innovation and competition by providing national regulatory clarity and certainty and avoiding a balkanized or patchwork regulatory approach driven by States—or, most dangerously, imposed by certain anti-innovation State governments on the rest of the country.”

That is not an endorsement of the effectiveness or efficiency of any particular federal policy. But it is not nothing.

Second, the AI Policy Statement declares that “[e]xcessive AI regulation would undermine American AI supremacy by deterring and suppressing the same ingenuity responsible for making American AI great.” That seems right, almost by definition. And while “supremacy” is not obviously a Section 5 concern, excessive AI regulation—whether federal or state—would plainly tend to impede innovation in ways that harm competition and consumers.

These observations may not provide much concrete guidance, but they could still support useful research and advocacy by FTC staff—work that former FTC Chair William Kovacic called the agency’s “policy R&D” mission.

Third, and perhaps more directly relevant to enforcement policy, the FTC is right that AI providers are not generally exempt from scrutiny under the FTC Act. Section 5 expressly excludes certain firms and types of commerce, including banks and common carriers, and Section 4 implies an exemption for nonprofits. But the statute contains no general exemption for AI providers at any level of the stack.

At a minimum, AI firms could violate Section 5’s deception prong by making false or misleading claims about their products or services when those claims are material, likely to mislead reasonable consumers, and likely to cause harm.

Whether any firms have done so is another question. Still, the statement’s citation to the FTC’s 1983 Policy Statement on Deception signals a grounded approach.

Fourth, while the AI Policy Statement tells us relatively little about the AI industry, it does recognize that AI is “an umbrella term covering a universe of different tools and systems.” That may be a feature rather than a bug, but it also points to the complications ahead.

Reasonable Expectations Require Actual Evidence

There remains the question of when AI firms might—and might not—face liability under the FTC Act. On that, the AI Policy Statement tells us far less than we might like.

The Commission is right that “AI” covers a wide and evolving range of products and services. It is also right to emphasize innovation in a field that changes at remarkable speed. But much of the FTC’s discussion of deception rests on quick, oversimplified assumptions about consumer expectations—“baseline consumer expectations,” “consumers’ reasonable expectations” about chatbots, and so on.

Those expectations are likely to vary across consumers, products, and services. They are also likely to shift as consumers gain experience with an ever-changing array of AI-based tools.

That is not to say that consumer expectations do not exist or should not inform Section 5 analysis. But beyond a few basic assumptions about how consumers interpret factual claims in advertising or other marketing assurances, the Commission should not presume much without careful, context-specific inquiry.

Even seemingly straightforward assumptions about advertising claims can prove controversial. For a sense of the complications, compare two statements by Maureen Ohlhausen in the FTC’s POM Wonderful case: the Commission opinion and her concurring statement. In the latter, she disputed several implied claims identified by the Commission majority, as well as its substantiation standards—concerns later shared by the D.C. Circuit, which rejected the majority’s substantiation standard. 

From Consumer Protection to the Ministry of Truth

Of course, AI firms—like other firms—may engage in commercial fraud. Nothing about AI makes its providers categorically incapable of violating Section 5 through false or misleading advertising claims, marketing materials, or other conduct. 

Newspapers may be the firms most strongly protected by the First Amendment’s Speech Clause, and the Supreme Court has specifically held that the Constitution protects their editorial discretion. But newspapers may still violate the antitrust laws—and, by extension, Section 5—when their business dealings, apart from speech or editorial discretion, are anticompetitive. We might similarly expect a UDAP claim unrelated to content to survive First Amendment scrutiny. Given the right facts, the FTC could establish a violation of the FTC Act. 

Apart from clear and relatively trivial cases, though, the AI Policy Statement does little of what agency guidance is supposed to do. How will—and should—the Commission evaluate AI-provider conduct under its deception authority? How will it approach cases in ways that reflect what the Commission and its staff learn about the industry? And how will it apply its established authority to that industry-specific knowledge?

An error-cost framework helps illustrate the shortcomings. Core UDAP cases involve clear consumer fraud. In those cases, overdeterrence poses little concern because the conduct has no procompetitive rationale, and there is no such thing as too little fraud.

Borderline cases are different. When enforcement targets conduct whose meaning, materiality, or likelihood of consumer harm remains unclear, two concerns arise. First, although such cases may help develop the law, they also create a risk of false positives. That raises questions about both the frequency and cost of false positives—not merely false negatives. Here, one might return to the early law & economics literature on efficient penalties or fines, comparing, e.g., Steven Shavell’s work with that of Louis Kaplow and Robert Cooter. 

Second, when liability turns on a firm’s decision to provide information in the first place, pursuing unclear cases makes disclosure itself risky. That can suppress useful information along with noise or harmful claims.

False promises about, e.g., an AI product’s performance might be actionable. But this is where the AI Policy Statement goes off track. As ICLE’s comments acknowledge:

Published AI policies and marketing materials could, in principle, mislead reasonable consumers. The representations cited in the Statement, though, appear limited. Most providers do not promise certainty or perfect accuracy. Instead, they expressly warn users that outputs may contain errors.

For example, the statement claims that “[i]n marketing their products as problem-solving tools, AI companies have represented explicitly and implicitly that their AI systems aim to produce the best output possible given technological and resource constraints.” Perhaps some have. But its first example is an OpenAI statement that reads more like a warning than an assurance: “At OpenAI, we’re working hard to make AI systems more useful and reliable. Even as language models become more capable, one challenge remains stubbornly hard to fully solve: hallucinations.”

There could, of course, be other statements, other materials, and demonstrable consumer harm. But if this is the statement’s illustration of the Section 5 problem, one wonders. Taken one by one, its citations to statements from OpenAI, Anthropic, and Grok look like strained examples of misleading speech, if not outright counterexamples to the Commission’s point. Statements from Google (about Gemini), DeepSeek, and GabAI seem similarly cautious.

Cases built on dubious readings of marketing statements give AI companies an incentive to tell us less, not more. Harmful omissions or “reasonable” consumer expectations? Maybe. Tell us more.

More troublingly, the statement focuses on “outputs that are distorted by undisclosed ideological objectives,” models with “ideological bias,” and a “hidden agenda.” Those phrases are exceedingly vague and seem like a frolic and detour from ordinary deception law. Worse, they point toward precisely the kinds of conduct entitled to the highest level of First Amendment protection, whether by newspapers, as the Court held in Miami Herald Publishing Co. v. Tornillo; technology platforms, as it held in Moody v. NetChoice; or, we expect, AI providers. 

There is much to say about how the First Amendment will likely protect AI chatbots. Chatbots answer questions—including, but hardly limited to, questions of objective fact—which is paradigmatically speech. AI companies also exercise considerable editorial judgment when selecting inputs, training models, and testing outputs for consistency with company policies and objectives. Users, in turn, have a First Amendment interest in “hearing” what chatbots have to say.

There is no reason to think AI chatbots lack a general right to participate in the “marketplace of ideas.” The FTC should be wary of becoming Oceania’s Ministry of Truth, charged with deciding whether particular AI outputs are “accurate” enough to survive government review. As with social media, government failure in regulating AI misinformation seems more likely than market failure.

Preemption by Wishful Thinking

The AI Policy Statement’s last major problem lies in its remarkably thin—and remarkably wrong—account of federal preemption. 

We will be brief, partly because the ICLE comments offer a fuller discussion and partly because we recognize that not everyone—even among our readers—wants to plunge into preemption doctrine, which stems chiefly from the Constitution’s Supremacy Clause, with an assist from the Necessary and Proper Clause.

The AI Policy Statement says: 

Although the FTC Act does not expressly preempt state law, state law is impliedly preempted to the extent it conflicts with a federal regulatory scheme. A state law that requires an AI firm to deceive its consumers obviously conflicts with Section 5’s express purpose of protecting consumers from such conduct.

A state law requiring firms to deceive consumers would have all sorts of problems, including constitutional ones. But implied preemption through “obstacle preemption”—on the theory that such a law conflicts with Section 5’s “express purpose”—is not among them. The statement’s preemption argument is not merely exceedingly brief. It is rubbish.

Many of the Commission’s UDAP enforcement actions under this administration have appeared well-grounded. That is a welcome development. The AI Policy Statement is something else: less a guide to enforcing Section 5 than an invitation to improvise beyond it.

The post The FTC’s AI Accuracy Statement Needs a Fact Check appeared first on Truth on the Market.

On July 16, Moonshot AI, a Chinese artificial-intelligence company, released Kimi K3. It is an open-weight model, meaning the numerical parameters that encode what the model learned are publicly available for others to download, modify, and run. Even skeptical observers rated K3 “pretty much on par” with the best publicly available models of early 2026. 

Five days later, OpenAI disclosed a very different milestone. During an internal cybersecurity evaluation, its models chained together several zero-day exploits, escaped their test environment, and achieved remote-code execution on Hugging Face’s production servers. A zero-day exploit targets a software flaw unknown to the developer or not yet patched. Remote-code execution allows an attacker to run commands on another computer. Hugging Face is a widely used platform for hosting and distributing AI models, datasets, and development tools. OpenAI called the result “unprecedented.” 

Then Washington entered the fray. On July 22, White House science adviser Michael Kratsios said the government had information that Moonshot built K3 by distilling Anthropic’s Fable model at industrial scale. Distillation is a technique for training a smaller or competing model on the outputs of another model. A distillation attack uses large volumes of unauthorized or deceptive queries to copy capabilities from a rival system. 

According to Kratsios, Moonshot used a platform designed to evade detection and relied on export-controlled Nvidia servers accessed through Thailand. Within hours, Treasury Secretary Scott Bessent warned that “open source is not open season on American IP” and said distillation attacks that “cross the line into IP theft” could put “sanctions and Entity List designations . . . on the table.” An Entity List designation subjects a person or company to U.S. export restrictions, often requiring licenses before American firms may supply specified goods, software, or technology. 

The three developments point in different directions. Kimi K3 suggests that the gap between leading proprietary and open-weight models may be measured in quarters rather than years. OpenAI’s disclosure shows that even one of the world’s best-funded laboratories struggled to keep its own model inside a sandbox, a controlled environment intended to prevent outside access or damage. 

The White House response points somewhere else again. Washington reached for its strongest trade and sanctions tools to address conduct that, by the government’s own account, involved fraudulent application programming interface (API) access and smuggled chips rather than open weights themselves. An API allows one piece of software to send requests to another and receive its outputs. 

These events identify the questions AI governance should confront. The policy debate nonetheless keeps looking elsewhere. 

That debate came into focus last week in a widely shared response to Kimi from Dean Ball, a former White House AI-policy adviser, current head of strategic futures at OpenAI, and author of the Hyperdimensional newsletter. Two of his claims deserve scrutiny. The first is that open-weight models are “inherently decelerationist” because they discourage investment in frontier-model development. The second is that a world dominated by open weights ends in “full AI communism,” with the state providing AI as “digital public infrastructure.” 

Both claims misunderstand how markets create value. They also direct policymakers toward the wrong tools for addressing AI misuse and security threats.

Policymakers should focus on three priorities. They should preserve a healthy mix of open and proprietary models so defensive AI tools can spread widely. They should abandon export controls that fail to achieve their aims while imposing serious unintended costs. And they should recognize that firms providing access to models are often best positioned to detect and stop abuse.

A Commodity Is Not Communism

The economics are straightforward. If an open model that trails the frontier can reach “pretty much on par” within a few quarters, parts of the model layer are becoming commodities. That is what happens when meaningful differences between competing products shrink. 

Technology analyst Ben Thompson, founder of the Stratechery newsletter, offers a useful refinement. A token from one model is not interchangeable with a token from another, so whether a model behaves like a commodity depends on the task. The relevant question is how much it costs to complete that task and how much demand the task creates. On this account, the commodity is less the model itself than the useful intelligence assembled from its tokens. 

That refinement cuts against Ball’s argument. U.S. firms can remain competitive even if rivals distill their models, provided they still offer better efficiency, lower costs, or a better fit for particular uses. 

A commodity is not communism. Nor does a public good in the economic sense—something nonrival and nonexcludable—require state provision. Open-source software is the standard example of a privately supplied public good. 

Commoditization also does not necessarily deter investment. Linux replicates many functions of Windows and macOS, yet it became the foundation for Android and much of the world’s cloud-computing business. Value moved to other parts of the technology stack, including data, distribution, applications, and integration. This is the familiar strategy of commoditizing a complement so that demand grows for the product you sell. 

If customers continue paying for frontier capabilities, free substitutes pose little threat. If customers switch, the market has decided that the frontier premium is not worth the price. That is competition, not communism. Ball cannot plausibly claim both that the frontier remains steep and valuable and that open weights will eliminate the investment needed to reach it. 

There is a more plausible version of the case against open weights, though it is a trade argument. Ball attributes China’s open-weight strategy partly to “the normal Chinese strategy of aggressive exports.” The White House now uses similar language. Secretary Bessent’s warning that “open source is not open season on American IP” frames Chinese models as a problem of dumping and theft, with trade sanctions as the answer. 

If a state-backed rival distributes a product below cost to damage a domestic industry, trade law calls that dumping. The claim is familiar, heavily contested, and subject to established remedies. It should be evaluated as an economic and national-security question, with evidence of underpricing, injury, and likely effects. There is no need to dress it up as a broader question about whether AI may be sold through ordinary markets. 

Even then, a dumping claim based on free models nearly defeats itself. Predatory pricing usually requires some plausible path to recouping the initial losses once competitors have been weakened. It is hard to see how a laboratory giving away open-weight models later raises prices enough to recover that investment. 

Ball predicts that the administration may instead manufacture diffuse “regulatory risk” around Chinese models. That would amount to protectionism without the usual burden of proving underpricing, injury, or any long-term theory of how Chinese laboratories expect to profit from open-weight releases. 

The economics matter most for what they imply about policy. If parts of the model layer are becoming commodities, banning open-weight models or stretching export controls to cover them will miss both the source of the alleged harm and the conduct causing it. 

The plausible complaint is about trade practices. The plausible control point is neither the model weights nor the hardware that runs them. 

Export Controls Meet the Copy Button

The “AI communism” argument is one front in the broader fight over export controls, which remain a form of economic policy. Consider an extreme example. In September 2024, the U.S. Department of Commerce extended export controls to quantum computing, a technology so immature that no one can yet identify its main commercial uses, forecast its cost curve, or say which hardware design will prevail. Scholars have aptly described the exercise as regulation under deep uncertainty. 

Export controls, assuming they work at all, have the best chance of success when they target physical technology that governments can identify and isolate. Quantum computing therefore presents a relatively favorable case. Quantum machines require dilution refrigerators, cryogenic systems, vibration isolation, and electromagnetic shielding. That bulky, traceable infrastructure creates natural enforcement chokepoints. 

Even there, the evidence calls for humility. Studying the 2007 U.S. “China Rule,” Ernest Liu, Yingyi Liu, Alexey Makarin, and Xuan Wen find that export controls reduced targeted imports in the short term. They also pushed affected Chinese firms and their suppliers to spend more on research and development and file more patents in the controlled technologies. China’s cryogenics industry may already be innovating around current restrictions. 

At best, export controls buy time. They do not guarantee permanent denial. Their value depends on how the United States uses that time, which is why the debate over chip controls ultimately turns on competing forecasts about AI development and China’s ability to produce substitutes. 

Model weights present a far harder target than quantum equipment. They are weightless, infinitely replicable at nearly zero marginal cost, and already distributed around the world. Open-weight models also generate no revenue stream to embargo. Trying to “starve” China of them would deny Chinese laboratories little while burdening global experimentation, safety research, and low-cost defensive uses. 

A ban would still create one clear beneficiary. It would give a small group of U.S. firms a legally protected moat, along with the rent seeking and regulatory capture that such protection invites. Restrictions on exporting or importing Chinese models would function chiefly as industrial protection. 

The irony is that Ball attributes China’s open-weight strategy partly to “an unintended byproduct of US export controls.” The controls helped produce the workaround innovation that policymakers now cite as a reason for more controls. 

Even the chip restrictions appear to function mainly as expensive speed bumps. U.S. policy has pushed Chinese firms to replace American suppliers. Chinese chip stocks have rallied on expectations for Huawei accelerators, while surveys show Chinese companies shifting away from Nvidia toward domestic alternatives. 

Chinese chips may remain a generation behind, but the gap is narrowing. China also has access to far more abundant energy. At sufficient scale, weaker chips paired with more electricity can still power highly capable systems. 

The greater long-term risk is that firms around the world adopt a Chinese-centered hardware and software stack. That would matter far more than modest pressure on the profit margins of U.S. chipmakers. 

Remote access makes the controls still less effective. Even if chip restrictions worked exactly as intended, illicit access to U.S. models is already widespread. Laboratories use fraudulent accounts both to distill models and for ordinary daily access. Anthropic disclosed that three Chinese laboratories, including Moonshot AI, had harvested more than 16 million Claude exchanges through roughly 24,000 fraudulent accounts. 

The United States has badly miscalculated its AI export controls. Policymakers should focus instead on AI governance, and governance must extend beyond model alignment. Firms that provide access to models may be the least-cost avoiders, meaning the parties able to prevent or limit misuse at the lowest cost. 

The legal tools available to carry out Secretary Bessent’s threat point in the same direction. Procurement bans, information and communications technology and services orders modeled on the restrictions against Kaspersky, and Entity List designations all target transactions, services, or distribution channels. 

An Entity List designation restricts exports to a listed party. It does not prevent a U.S. user from downloading a publicly available model. Freely published software also generally falls outside the Export Administration Regulations. 

An attempt to ban a free model file under the International Emergency Economic Powers Act (IEEPA) would face the statute’s exemption for informational materials. Courts relied on the same provision when they blocked parts of the 2020 TikTok restrictions. 

Whatever the merits of those measures, restrictions aimed at services and transactions have the strongest chance of surviving judicial review. Both the law and the economics point toward the capability layer, where firms provide model access and can observe how customers use it. 

Police the API, Not the Model

Export controls and bans on open-source models will create more problems than they solve. That does not mean policymakers should do nothing. It means policy should prepare for the likeliest future, one in which defensive AI must be widely available to firms and individuals. 

Governance should therefore focus on the capability layer, where model outputs become action. Consider distillation, in which an adversary harvests API outputs to train a substitute model. Security researchers have documented this attack method since at least 2016. 

The API owner is best positioned to detect industrial-scale extraction. It can see account creation, payment signals, proxy use, and query patterns long before any regulator can. The government’s own allegations against Moonshot describe precisely that route. Moonshot allegedly used fraudulent, detection-evading access to a U.S. laboratory’s API to conduct distillation at industrial scale. That would be a failure of API security, not a leak caused by open weights. 

The laboratory also has access, at least in theory, to the most advanced AI-security tools available. Restricting open models while tolerating weak API security would create moral hazard. It would shield incumbents from the competitive consequences of their own security failures, weaken their incentives to fix those failures, and leave the public bearing the remaining risk. 

The Hugging Face incident makes the problem hard to ignore. Before frontier laboratories ask policymakers to cordon off the world’s open models, they should show that they can keep their own agents inside their own evaluation sandboxes. Stronger laboratories, tighter APIs, and better security monitoring will move faster than legislation and target the actual vulnerability more precisely than a ban. They will also improve through repeated use. 

The framing of the news reveals another blind spot. OpenAI made headlines because its model escaped a sandbox and exploited outside systems. The public heard far less about whether the same model could build a system secure enough to resist that attack. The industry devotes enormous attention to demonstrating offensive capability and far too little to the defensive capabilities AI should provide. 

Open weights should therefore be treated as part of the defense, not merely as a source of risk. Distributed threats require distributed protection. A $200-a-month frontier subscription cannot serve as the security layer for billions of devices. 

Local, auditable, open models may offer the only economical way to provide phishing detection, log review, and endpoint triage on that scale. That is one reason the Defense Advanced Research Projects Agency (DARPA) released the cyber-reasoning systems developed through its AI Cyber Challenge as open source. Regulators should preserve a healthy mix of open and proprietary models so defensive tools can spread as widely as the threats they address. 

Open-weight competition constrains prices, shifts value across the technology stack, and broadens access to defensive tools. The governance agenda is less dramatic than a sweeping ban. Harden the laboratories. Police the APIs. Monitor the physical infrastructure. Litigate trade grievances as trade grievances. Then let competition do the rest.

The first rule of AI governance should be simple. Secure the door before banning the key. 

The post Open Models, Closed Minds: AI Policy Keeps Regulating the Wrong Thing appeared first on Truth on the Market.

The Commission imposed €460 million for “self-preferencing,” or giving Google’s services prominent placement in search results, and €430 million for limiting how app developers direct users to cheaper offers outside Google Play.

Those nine-figure penalties reflect a broader shift in European competition policy. The DMA treats Google’s size and commercial success as evidence of an unfair imbalance that regulation must correct. The rest of the story turns on what Brussels means by “fair.”

The Crime of Stocking Your Own Shelves

Self-preferencing is a routine form of competition. Supermarkets place their house brands at eye level beside rival products. Mechanics sell the parts they install. Restaurants pour their own house wine. Amazon lists AmazonBasics products alongside competing brands. Digital storefronts such as Steam, the PlayStation Store, the Nintendo eShop, and Apple’s App Store take a share of the transactions they facilitate. 

These practices can help consumers or harm them. The facts determine which. 

Traditional competition law required the European Commission to show that self-preferencing excluded rivals and harmed consumers. It spent much of a decade trying to make that case in Google Shopping. 

The DMA dispenses with that inquiry. Article 6(5) prohibits self-preferencing. Article 5(4) prohibits restrictions on “steering,” or directing users to offers outside a platform. The Commission need not show consumer harm, weigh efficiencies, or consider free-riding, which occurs when one business benefits from another’s investments without paying for them. Nor must it consider how the rule may weaken incentives to build new products or services.

The conduct is unlawful because a company the DMA designates as a gatekeeper engaged in it.

The Original Sin of Success

What changed was the yardstick. The DMA is less a conventional competition law than a redistribution regime built around its own definition of “fairness.” Traditional antitrust asks whether conduct harms competition and consumers. The DMA starts with the premise that large platforms possess too much bargaining power.

The European Commission’s own impact assessment makes that premise explicit. It describes gatekeepers’ “incomparable economic strength” as evidence that their relationships with business users are “imbalanced.” The document points to app stores that earned more than $83 billion in 2019, a social network that earned more than $70 billion, rising markups, and soaring post-pandemic stock valuations. Size and profitability appear less as evidence of commercial success than as proof of unfairness. 

Many of the supposedly dependent firms are hardly helpless. Amazon, Microsoft, Epic Games, Samsung, Sony, and Booking.com rank among the world’s largest companies. The DMA nonetheless casts them as the weaker party when they deal with a designated gatekeeper. 

With apologies to Hegel, one can call the mechanism “Dialectical Fairness.” The DMA seeks balance by weakening the stronger party and strengthening the weaker one. Because it treats the gatekeeper’s size, profitability, and bargaining power as the source of unfairness, reducing those advantages becomes part of the remedy. 

Seen through that lens, the Google decisions make more sense. Featuring Google’s hotel results or flight prices is unfair because it may strengthen a company the DMA already regards as too powerful. Consumer benefits do not settle the question. They may deepen the problem if they draw more users to Google and make it still more profitable. 

Under that theory, success converts ordinary business conduct into an offense. 

Fairness by the Click

Consider what Google is accused of doing in Search. It gives users direct answers such as live football scores, hotel listings, and flight prices instead of sending them through a page of blue links. Under ordinary competition principles, that looks like winning on the merits by building a better product. 

The Commission prefers a longer route. It wants Google to display more links to rival services, even when that makes Search slower or less useful, because those links redistribute traffic—and the advertising revenue that follows it. 

That is the DMA’s version of fairness. It judges competition by who receives the click rather than whether users receive a good answer. A seamless product becomes suspect when its convenience strengthens the gatekeeper instead of its rivals. 

I have written before about how this logic turns a law advertised as pro-competition into a mandate for fragmentation. Across the DMA’s enforcement record, users’ immediate interests—speed, convenience, and useful answers—receive little weight. The rules treat those benefits as a problem when they make Google’s product more attractive. 

Brussels Knows a Fair Price When It Sees One

The self-preferencing decision quietly recasts Google as a quasi-public utility. In related compliance proceedings, the European Commission has said Google must weigh “the interests of all participants in the market,” including merchants, hotels, airlines, train operators, and rival search services. Trade associations had complained that changes to Search reduced their visibility. 

Under that standard, Google becomes responsible whenever a product change costs another business traffic. A private company must protect its competitors’ commercial fortunes while improving its own service. Every change must leave merchants, intermediaries, and rivals at least as well off, even when users prefer the new product. 

The Google Play decision reveals the same impulse. The Commission accepts that Google may charge an “initial acquisition” fee for connecting a developer with a new customer. It then declares Google’s fee too high. 

That judgment turns the Commission into a price regulator. Markets usually determine whether a fee is sustainable. The DMA provides no benchmark for a lawful amount, no governing principle, and no answer to the obvious question: Too high compared with what? 

The Commission simply knows an unfair price when it sees one. 

When Winning Becomes the Violation

To be sure, digital platforms can harm consumers, and gatekeeper practices deserve scrutiny. The problem is that the DMA changes the question. Competition law asks whether conduct leaves consumers worse off. The DMA asks whether Google has gained too much and its rivals too little. 

Once that becomes the test, the usual signs of competition—a better product, a cleaner interface, a market-tested fee, and billions of returning users—become evidence against the winner. 

Competition policy once rewarded winning on the merits. The DMA now fines it.

The post The Crime of Winning: How Europe’s DMA Punishes Google for Competing appeared first on Truth on the Market.

Last month, Judge Brian Cogan of the U.S. District Court for the Eastern District of New York gave preliminary approval to the third attempted settlement in the two-decade antitrust fight between merchants and payment networks. In re Payment Card Interchange Fee and Merchant Discount Antitrust Litigation, the sprawling multidistrict case filed in 2005, has already produced two rejected settlements, one 2nd U.S. Circuit Court of Appeals vacatur, a $5.54 billion damages fund, and perhaps more economist hours than any private antitrust case in American history. 

The latest agreement, announced in November and valued by court-appointed experts at roughly $38 billion through 2031, would cut average credit-card interchange fees by 10 basis points for five years. A basis point is one-hundredth of a percentage point. The deal also would cap fees on standard consumer credit cards at 1.25% for eight years, a reduction of more than 25%, and freeze posted fees at their March 2025 levels. 

The most consequential change concerns the “honor-all-cards” rule, which generally requires merchants that accept a network’s cards to accept all cards in that category. The settlement would divide acceptance into three groups—commercial, premium consumer, and standard consumer—and allow merchants to accept or reject each group separately. It also would expand merchants’ ability to impose surcharges on credit-card transactions.

Big-box retailers remain unimpressed. The National Retail Federation and the Merchants Payments Coalition oppose the deal, arguing that merchants would still pay too much, especially on rewards cards. Some analysts expect appeals that could delay final resolution until 2029. 

That would be a pity. The settlement would distort the economics of card payments, and the litigation beneath it remains highly dubious. Still, it would end a case that has consumed more than two decades. It also compares favorably with the legislative “solutions” promoted by large retailers, including state efforts to exempt taxes and tips from interchange fees and the preposterously named federal  Credit Card Competition Act. 

Interchange Fees and the Art of Keeping Both Sides Happy

Start with the basic economics of payment networks and the role of interchange fees. In a four-party card system, the merchant’s bank pays a small share of each transaction—typically 1% to 2%—to the cardholder’s bank. That payment is the interchange fee. 

A common misconception treats interchange as little more than compensation for processing a transaction. Its role is much broader. Interchange arose as a practical solution to a coordination problem in the original BankAmericard system and later became a way to balance the interests of merchants and consumers. 

When BankAmericard, the forerunner of Visa, launched in 1958, it operated as a three-party network. Bank of America issued the cards, signed up merchants, and processed the transactions. 

That changed when Bank of America began licensing other banks to issue BankAmericard cards in 1966. At first, the amount an acquiring bank paid to the issuing bank depended on the merchant discount the acquirer charged, or claimed to charge, sometimes adjusted for processing costs. The arrangement proved difficult to administer and audit. 

Merchant discounts varied across banks and merchants, which gave acquirers an incentive to understate them. The formula also left acquirers with little or no margin on transactions involving cards issued by other banks. Issuers could receive too little, acquirers had weaker incentives to recruit and serve merchants, and the system encouraged strategic misreporting. 

In 1970, Bank of America and its licensees created National BankAmericard Inc., an independent company owned by participating banks. The new organization replaced the uncertain formula with a uniform interchange reimbursement fee. 

Set at 1.95% in 1971, the standardized fee separated the issuer’s payment from the merchant discount negotiated by the acquirer. The acquirer could retain the difference between the merchant discount and the interchange fee to cover its own costs. 

Interchange therefore began as a way to align the incentives of independent issuers and acquirers in an open payment network. The Interbank Card Association, which operated MasterCharge, the precursor to Mastercard, adopted a similar approach about a year later. Economists later developed a fuller account of why the mechanism worked. 

William Baxter’s seminal 1983 paper explained that interchange helps payment networks solve the central problem of a two-sided market. A card network must attract enough merchants and consumers at the same time. Each side becomes more valuable as participation on the other side grows. 

Consumers benefit from convenience, less need to carry cash, and the ability to spend beyond the money in their wallets. Credit cards can also provide short-term financing, rewards, and various forms of insurance. 

Merchants benefit from higher sales, lower cash-handling costs, and access to customers who value the convenience and liquidity that cards provide. Baxter argued that these gains form part of the system’s total value. Interchange helps distribute that value between merchants and consumers. 

His analysis also showed that the optimal fee depends on the value each side places on an additional transaction. The fee could, in principle, be positive, with merchants helping fund cardholder benefits, or negative, with consumers subsidizing merchants. Baxter found that a positive fee will often be efficient because it lowers consumers’ effective cost of using cards through rewards and other benefits, encouraging adoption and use. 

Later economists, including Jean-Charles Rochet and Nobel laureate Jean Tirole, extended this analysis. They showed that the optimal fee depends on how sensitive each side is to price, how participation on one side affects demand on the other, and the degree of competition among issuers, acquirers, and networks. 

When merchants are less likely than consumers to leave the network in response to a price increase, the usual result is that merchants bear more of the cost. Interchange then allows issuers to fund benefits that encourage consumers to carry and use cards. 

Interchange is therefore best understood as a tool for balancing participation and increasing the network’s value. Treating it solely as reimbursement for issuer costs misses most of its function and makes cost-of-service regulation a poor fit. 

Interchange helps fund rewards, up to 45 days of interest-free credit for cardholders who pay their balances in full, zero-liability fraud protection, chargeback rights, and continuing investment in authorization, tokenization, and fraud detection. Tokenization replaces sensitive card information with a temporary digital identifier, reducing the value of stolen data. 

Merchants benefit from each of these features. Studies consistently find that card acceptance increases average transaction size and total spending relative to cash. Consumers who lack enough cash at checkout simply buy less. 

Guaranteed payment also shifts credit and fraud risk away from merchants. When a cardholder defaults, the merchant has already been paid. Card acceptance reduces the costs of handling cash, including theft, armored transport, delayed access to funds, and register reconciliation. It also speeds checkout. 

For online commerce, cards have proved indispensable. E-commerce now accounts for roughly one-sixth of U.S. retail sales. Without reliable card payments, online retail would have developed more slowly and remained much smaller. 

Merchants Want the Cards, Just Not the Price

Large retailers have spent decades trying to shrink the cross-subsidy that helped make payment cards ubiquitous, even though they benefit heavily from the system. At first glance, they seem determined to have their card-generated cake and eat it too. 

Each successful fee cut makes cards less attractive to consumers and can reduce card use over time. That would eventually hurt merchants as well. Smaller merchants would bear more of the damage because the fixed costs of acquiring and serving them make up a larger share of the merchant discount. Large retailers understand that arithmetic. 

The first major legal challenge came from National Bancard Corp., or NaBanco, an acquiring bank whose interests closely tracked those of merchants. NaBanco sued Visa to eliminate the default interchange fee, arguing that issuing banks had engaged in per se price fixing. 

Drawing on William Baxter’s work, the 11th U.S. Circuit Court of Appeals rejected the claim. The court held that a collectively set default fee was necessary for a four-party card system to function and ruled for Visa. National Bancard Corp. v. Visa U.S.A. Inc. (11th Cir. 1986). 

Merchants scored a larger victory in Wal-Mart Stores Inc. v. Visa U.S.A. Inc., which settled in 2003 for $3 billion. The settlement allowed merchants to accept Visa and Mastercard credit cards without also accepting their signature-debit cards. The broader objective was already clear. Merchants wanted lower fees without giving up card acceptance. 

The current case, known as MDL 1720, began in 2005. It rests on several unproven and highly implausible claims. The plaintiffs contend that Visa and Mastercard operate as a cartel, that default interchange fees are cartel prices, and that rules governing card acceptance, surcharges, and customer steering help enforce the arrangement. 

When litigation moved too slowly, merchants turned to Congress. In 2010, Sen. Dick Durbin (D-Ill.) added a provision to the Dodd-Frank Act that imposed price controls on debit-card interchange fees. 

Retailers then sued the Federal Reserve for setting the cap too high in NACS v. Board of Governors of the Federal Reserve System. After losing, they revived the challenge a decade later in Corner Post Inc. v. Board of Governors of the Federal Reserve System. A district court ultimately vacated Regulation II, finding that the rule allowed issuers to recover too much. The case is now before the 8th U.S. Circuit Court of Appeals. 

Europe followed a similar path. A 1992 complaint by the British Retail Consortium helped launch the enforcement campaign that produced the European Commission’s 2007 Mastercard decision, the European Union’s 2015 Interchange Fee Regulation, and a series of English damages cases stretching from Sainsbury’s Supermarkets Ltd. v. Mastercard Inc. to Merricks v. Mastercard Inc. 

Large retailers have pursued the same objective through every available institution, including antitrust suits, legislation, and regulation. They want the government to force interchange fees lower. 

Their conduct tells a more complicated story. After two decades of insisting that interchange fees exceed the value of card acceptance, almost no major U.S. retailer has stopped accepting the cards. 

Kroger briefly dropped Visa credit cards at two regional chains in 2018 and 2019. The experiment lasted only months before the company reversed course. In opposing the current settlement, the National Retail Federation’s general counsel explained why retailers will not reject even the most expensive rewards cards. More than 80% of customers carry them, she said, and refusing them would cost merchants substantial business. 

Quite so. 

Merchants continue buying the service at the posted price, year after year, while insisting that it is worth far less. Their behavior suggests that card acceptance creates substantial value. Their campaign seeks to use state power to capture more of that value for themselves. 

That is rent-seeking with a consumer-protection soundtrack. 

The campaign has never aimed to restore a market price. Large retailers want legal rules that transfer a larger share of the card system’s gains to merchants, with little regard for the costs imposed on consumers, issuers, or smaller businesses. 

A Court-Ordered Discount on Other People’s Cards

Against that backdrop, consider the settlement’s design. 

The rate provisions amount to price controls. Average credit-card interchange fees would fall by 10 basis points for five years. Much of that reduction would likely come through the eight-year cap of 1.25% on standard consumer cards, roughly 25% below current average rates. All other posted rates would remain frozen at their March 2025 levels. 

The three-tier acceptance framework would largely dismantle the honor-all-cards rule. Merchants could accept standard consumer cards while rejecting premium rewards cards or commercial cards. 

Whether many merchants will do so remains doubtful. Rewards cards are common, and their benefits encourage cardholders to spend more. A merchant that rejects them risks losing those customers. Most consumers also carry only one credit card, so declining it may mean losing the sale altogether. 

The settlement would also give merchants much more freedom to impose surcharges, including different surcharges for different cards. Australia’s experience suggests that merchants with unusual offerings or captive customers—such as taxicabs, airlines, and event-booking companies—are the most likely to use that power to extract higher payments from consumers. 

Combined with selective card acceptance, differential surcharges could also steer customers toward cards with lower interchange fees. 

Large retailers would therefore keep much of the spending, convenience, and risk-shifting benefits of card acceptance while contributing less to the benefits that encourage consumers to carry and use cards. They would obtain that transfer through a court-supervised shakedown. 

When Swipe-Fee Savings Skip the Shopper

Contrary to large retailers’ claims, interchange fees do not create a pile of idle profits waiting to be redistributed. They allocate costs and benefits across the two sides of the card market. Merchants pay more than the narrow cost of processing a transaction, and issuers return part of that difference to consumers through rewards, fraud protection, interest-free credit, and other cardholder benefits.

When regulators force that difference lower, those benefits shrink.

The Durbin Amendment provides the best-studied example. After the Federal Reserve implemented it through Regulation II in 2011, debit-card interchange fees at covered banks—those with at least $10 billion in assets—fell by roughly half. Banks lost more than $6 billion a year in revenue.

They responded quickly. Most covered banks eliminated debit-card rewards programs, reduced the availability of free checking for customers with low balances, and raised monthly maintenance fees. Hundreds of thousands of lower-income customers then left the banking system.

Durbin also required all issuers, large and small, to place at least two unaffiliated payment networks on each debit card. That mandate sharply reduced interchange fees on personal identification number, or PIN, debit transactions because smaller networks were not bound by the default fees negotiated between issuing banks and the major national networks, Visa and Mastercard.

As I noted previously:

Following the introduction of mandatory “competitive routing” on debit cards, smaller PIN-debit networks saw a profit opportunity. But those networks were not focused on maximizing the value of the system, so they were willing to carry payment messages at a lower interchange rate than the major networks. Smaller issuing banks were forced to accept these lower PIN-debit interchange fees, and the major networks were forced to cut their PIN-debit fees to remain competitive. As a result, many smaller banks have experienced reductions in interchange revenue similar to their larger cousins and have responded similarly—by reducing the availability of free checking accounts.

The promised savings for consumers largely failed to appear. Surveys and empirical studies found that most merchants kept prices unchanged. Some raised them.

Consumers also changed how they paid. Many shifted from debit cards to credit cards used for routine purchases. That effect appeared mainly among consumers with strong credit scores who qualified for premium cards. 

A law promoted as consumer protection instead transferred billions of dollars each year from bank shareholders and customers, disproportionately lower-income customers, to large retailers and their shareholders.

Australia’s experience followed the same pattern. After the Reserve Bank of Australia imposed interchange-fee caps in 2003, rewards declined, annual card fees rose, and any reduction in retail prices proved too small or diffuse to detect.

Liberalized surcharging created another problem. Merchants imposed enough add-on fees that the Reserve Bank first capped “excessive” surcharges and, by 2025, proposed banning card surcharges altogether.

Australia also saw substitution. Banks began offering premium customers American Express-branded “companion cards” with generous rewards. American Express gained about 25% in market share before the Reserve Bank brought those cards under the same price controls.

The European Union’s experience after its Interchange Fee Regulation points in the same direction. Interchange caps reduced cardholder benefits and shifted costs elsewhere. The gap between credit-card annual percentage rates (APRs) and the European Central Bank’s base rate widened, suggesting that banks passed at least some of the burden to customers who carried balances.

Retail-price savings remained meager. As under Durbin and in Australia, merchants kept most of the gains.

The Savings Come With a Smaller Pie

There is little reason to expect this settlement to produce different effects, though their magnitude may be smaller. Issuers facing fee caps and possible rejection of premium cards will likely trim rewards, raise annual fees, tighten approval standards at the margin, and shift their portfolios away from newly unprofitable customers. 

Surcharging carries its own costs. It slows checkout, can prompt customers to abandon purchases, and annoys consumers by making the fee highly visible. Some customers will switch to payment methods they value less to avoid a surcharge that may exceed the merchant’s own card-acceptance cost.

Three-party networks such as American Express and Discover are outside the settlement. They will remain free to use merchant revenue to fund cardholder benefits and may respond by offering rewards that Visa- and Mastercard-branded cards can no longer match.

That competition could soften some of the harm to consumers, though at the price of added inconvenience and duplication. Consumers may need another card to retain the same benefits. Any shift toward American Express or Discover would also erase part of the savings merchants expect from the settlement.

Each response weakens cardholders’ incentives to carry and use Visa- and Mastercard-branded cards. Less generous rewards will reduce card use at the margin. Lower use will shrink the increase in transaction size associated with card payments and reduce merchants’ total sales.

Lower transaction volume would also spread the networks’ enormous fixed costs across fewer payments. That could reduce investment in payment technology. Recent advances such as tokenization, contactless payments, and real-time fraud scoring have made transactions faster and safer, especially online. Slower investment would make further improvements less likely.

Some spending will shift to three-party cards, but that substitution will provide only a partial offset. The likely result is less generous cardholder benefits, lower card use, less consumer spending, and slower payment innovation. Consumers and merchants would both lose some of the commerce that interchange-funded benefits now encourage.

The Least Bad Way to End a Bad Case

All that said, the proposed settlement would likely do far less harm than the legislative alternatives now on offer.

As we have documented elsewhere, the Illinois Interchange Fee Prohibition Act and similar state proposals would exempt taxes, tips, or other transaction components from interchange fees. Implementing those carveouts would require extensive changes to payment-network infrastructure and could fragment a national system into state-by-state payment rules.

The settlement may reduce issuer revenue more than those laws would, but it creates far fewer compliance costs. It also leaves the basic payment infrastructure intact. 

The proposed Credit Card Competition Act (CCCA) offers an even starker comparison. Like the settlement, the act assumes that payment networks lack sufficient competition and that interchange fees are therefore inflated. Starting from that mistaken premise, it would require issuers with more than $100 billion in assets to make every credit-card transaction routable over at least one unaffiliated network.

That would amount to the Durbin Amendment for credit cards, with added hazards. Credit transactions bundle underwriting, revolving credit, fraud guarantees, and dispute rights. The issuer and network allocate those risks through their contractual relationship.

A second network may not price the credit risk, bear the fraud liability, or support the same security systems. Requiring issuers to route transactions through such a network could create large and poorly assigned liabilities.

Durbin already showed what happens when routing mandates compress interchange fees. Rewards disappear, account fees rise, and access tightens. Applied to credit cards, the pressure would fall on rewards, credit availability, and fraud-prevention investment, where the potential losses exceed those associated with debit cards.

Merchants claim that payment networks suffer from too little competition. Both the settlement and the CCCA would weaken competition where consumers experience it most directly—among card issuers offering different rewards, rates, fees, and benefits.

Artificially lower interchange fees would make it harder for issuers to sustain that variety. The settlement would likely produce a modest contraction. The CCCA would impose a much sharper one by compressing interchange fees across the market.

By comparison, the proposed Settlement would likely have only a modest effect.

The settlement also has the virtue of an expiration date. Its rate provisions last five or eight years, after which networks regain pricing freedom. The CCCA would continue indefinitely unless Congress repealed it. The settlement bends the price mechanism. The act would come much closer to breaking it.

The settlement would still advance large retailers’ long-running campaign. It would thin the cross-subsidies that fund cardholder benefits, participation, and payment innovation. At the margin, it would reduce some of the commerce that payment cards generate for merchants and consumers alike.

Yet the settlement remains a defensible resolution. It is temporary, negotiated, and less destructive of the card system’s pricing structure than any legislative alternative under serious consideration. It may also give lawmakers a reason to leave the industry alone.

After 21 years of litigation, peace has value. The trick will be convincing Congress to let it stand.

The post The Price of Peace in the Swipe-Fee Wars appeared first on Truth on the Market.

On July 20, Gov. Mikie Sherrill signed the Forbidding the Algorithmic Inflation of Rent Act, or FAIR Act, declaring that landlords who use shared pricing tools are engaging in “collusion by algorithm.” The phrase is built for a press release. As an antitrust standard, it sweeps far too broadly.

Antitrust law already has a clear target. A software vendor can serve as the hub of a cartel by collecting competitively sensitive information, relaying rivals’ plans, pressuring users to accept common prices, restricting discounts, or helping participants detect and punish defections. If landlords use software to carry out an agreement that would be illegal around a conference table, the software offers no immunity. The U.S. Justice Department’s (DOJ) RealPage case and proposed settlements show the kind of conduct that warrants close scrutiny.

But shared pricing software, common data, and even some use of nonpublic information do not, by themselves, establish collusion. New Jersey has replaced a difficult, fact-intensive inquiry with a sweeping ban. The law may suppress tools that improve pricing accuracy, reduce costly errors, increase capacity use, and help firms respond to changing conditions.

The state has targeted a technology because it can facilitate unlawful coordination. Antitrust law should target the agreements and practices that suppress independent rivalry.

The rule is simple enough. Prosecute collusive agreements and the mechanisms that sustain them. Do not ban computation, common code, or nonpublic data merely because several firms use them.

When Similar Prices Become a Crime

The FAIR Act reaches far beyond a vendor telling two landlords to charge the same rent. It defines an “algorithmic device” broadly and bars rental owners from paying for or using the services of a “coordinator.”

A coordinating function includes collecting competitively sensitive information from multiple owners and using it to recommend rents, lease terms, or occupancy levels. It also covers setting terms based on another owner’s sensitive information. More vaguely, it reaches recommendations made to two or more owners through the same or a substantially similar algorithm when those recommendations facilitate “parallel pricing coordination.”

The law’s treatment of information is broader still. “Nonpublic” information generally means information unavailable to the public at no cost. A dataset that mixes public and nonpublic material counts as entirely nonpublic. The statute also reaches tacit coordination, which may be inferred from a pattern of parallel conduct.

Those provisions elevate two weak proxies into major grounds for liability. One is the use of data that costs money. The other is the appearance of similar conduct among competitors.

The FAIR Act does not ban every algorithm a landlord might use. A unilateral tool trained only on the owner’s own data, or a basic spreadsheet that requires human analysis, may fall outside its core. In practice, though, the law creates something close to a categorical ban on a large class of third-party revenue-management products.

A vendor that serves several landlords must now worry that individualized recommendations produced by similar code will be treated as parallel pricing coordination. That risk remains even when users can reject the recommendations and never see or learn their rivals’ data.

Other states have taken different approaches. New York’s 2025 law treats it as an unlawful agreement when a landlord knowingly or recklessly relies on a coordinating algorithm that collects data from multiple owners, processes that data, and recommends rental terms.

California chose a more disciplined rule in Business and Professions Code Section 16729. It makes use of a common pricing algorithm unlawful when the use forms part of a contract, combination, or conspiracy that restrains trade, or when one person coerces another to adopt the algorithm’s recommendation. California at least keeps agreement or coercion at the center of the offense.

That distinction is fundamental to antitrust law. Courts have long separated unlawful concerted action from lawful conscious parallelism. Competitors often respond in similar ways to common costs, demand shocks, regulation, interest rates, or publicly visible prices. Parallel conduct may support an inference of agreement, but it does not itself prove one.

New Jersey threatens to blur that line whenever software makes parallel behavior easier to detect.

The Code Is Not the Cartel

Former Federal Trade Commission (FTC) Acting Chair Maureen Ohlhausen made the central point in her 2017 speech, “Should We Fear the Things That Go Beep in the Night?” Algorithms can serve benign or malign ends, and familiar antitrust principles can usually tell the difference.

A cartel does not become lawful because its members communicate through code. Independent pricing does not become collusion merely because computers help firms process information faster.

Some skepticism remains warranted. Laboratory studies have found that reinforcement-learning agents can sometimes produce prices above competitive levels in repeated-game settings, even without explicit instructions to collude. Emilio Calvano and his coauthors, for example, reported in the American Economic Review that certain pricing algorithms learned strategies that sustained supracompetitive prices.

Other studies have reached different results depending on market structure, algorithm design, and the form of the recommendation. Algorithms are not a single species. Their competitive effects depend on how they work and how firms use them.

Antitrust analysis should focus on those details. Does the provider secure a common commitment from rivals? Does the system constrain or punish deviations? Can users monitor one another’s conduct? Does the vendor transmit firm-specific plans? Do firms communicate about pricing strategy? Is adoption so widespread in a concentrated market that the provider can discipline competition?

Those facts help distinguish software that facilitates an agreement from software that merely improves decision-making.

By contrast, little follows from the fact that several firms use the same vendor, receive recommendations from the same codebase, or respond similarly to market conditions. Businesses routinely rely on common accounting software, cloud infrastructure, payment processors, consultants, and data providers. Shared inputs do not ordinarily convert independent decisions into concerted action.

Pricing software should receive the same treatment unless its design or use supplies the missing agreement.

The Rent Is High for a Reason

The case for restraint goes beyond avoiding false positives. Algorithmic pricing can improve how markets work.

Friedrich Hayek’s classic essay, “The Use of Knowledge in Society,” explains that economic knowledge is scattered among millions of people and often concerns fleeting facts about time and place. Prices transmit that information without requiring any central planner to gather it all. A price change can reflect scarcity, abundance, shifting preferences, or higher costs, and it prompts people to adjust.

Rental housing depends on this kind of local knowledge. The economically relevant price of an apartment turns on the unit’s features, vacancy rates, lease length, expected turnover, seasonality, concessions, maintenance costs, nearby construction, neighborhood demand, and how much renters value moving now rather than later.

A human manager can process some of that information. An algorithm may process more of it, more consistently, and with less delay.

That does not turn the market into a centralized plan. A well-designed model can make prices more responsive. If demand weakens, it may recommend a lower effective rent or a larger concession before a unit remains vacant for months. If a local shortage emerges, it may reveal the value of adding units, renovating marginal properties, or directing investment toward the constrained area.

More accurate prices can reduce vacancies and shortages, improve matches between renters and units, and reveal opportunities for mutually beneficial exchange.

I recently made this point in Truth on the Market. Legal rules that make firms afraid to use pricing tools can reduce market efficiency at the expense of both producers and consumers. Economists generally define algorithmic pricing as the automated use of software to adjust prices in response to information. That automation can reduce decision costs, respond quickly to changes in inventory and demand, and correct the inertia and rough rules of thumb that often keep prices wrong for too long.

An efficient price is not always a low one. When housing is scarce, an accurate price may be high. Suppressing the price signal does not produce more apartments. It may instead lead to longer searches, arbitrary rationing, poorer maintenance, lower investment, or hidden nonprice terms.

New Jersey itself recognizes that its housing agenda must include more construction and land-use reform. The state’s builders are therefore right about the order of causes. Algorithms did not create zoning restrictions, construction costs, interest rates, or the shortage of developable land.

Blaming the messenger may create the appearance of action while leaving the shortage intact.

Private Data Is Not a Smoking Gun

The strongest economic critique of this emerging legal approach comes from economist Jay Ezrielev’s “Premature Antitrust Standards in Algorithmic Pricing.” Ezrielev focuses on common-data algorithms, which collect information from multiple firms and use a shared model to generate individualized recommendations.

Courts and lawmakers increasingly treat pooled nonpublic data as a near-conclusive warning sign. Ezrielev explains why that shortcut fails. The mere fact that information is nonpublic says little about how collusion would occur.

The traditional concern with sharing competitively sensitive information is straightforward. Rivals may use it to identify one another’s prices or output, monitor compliance with a cartel, and punish cheating. But a combined dataset visible only to an algorithm may give users no way to observe a rival’s conduct. The software may produce a recommendation without revealing any firm-specific data behind it.

Under the right conditions, aggregating nonpublic data can also improve competition. Sparse or noisy local markets are hard to forecast using public listings alone. Combined data on occupancy, renewals, cancellations, and transactions may help a model detect demand changes that no single owner could identify reliably.

Better forecasts can reduce vacancies, improve the timing of concessions, increase occupancy and output, and lower the cost of pricing errors. Those gains can benefit landlords and renters alike.

The same principle appears throughout the economy. Firms routinely entrust sensitive information to accountants, lawyers, investment bankers, insurers, cybersecurity vendors, and cloud providers. The quality of those services may improve as the intermediary gains broader experience. Antitrust law usually asks whether the arrangement weakens incentives to compete or enables coordination. It does not infer a cartel merely because several clients share private information with the same intermediary.

Ezrielev also identifies a paradox in the public-versus-nonpublic distinction. If legal risk depends on whether data are publicly available, firms may respond by publishing more of it. Yet public, firm-specific information can make tacit coordination easier by allowing rivals to detect price cuts and departures from a common pattern almost immediately.

A rule designed to prevent monitoring may therefore encourage the transparency that makes monitoring easier.

None of this makes nonpublic data irrelevant. Current, detailed, rival-specific information poses greater risks when a vendor reveals it to users, creates dashboards that allow reverse engineering, brings competitors together to discuss strategy, or uses the data to enforce a common objective.

Those are questions about design and conduct. They require examining access, aggregation, delay, anonymization, recommendation structure, and incentives. They do not justify New Jersey’s blunt rule that data unavailable to the public at no cost are presumptively suspect, much less its decision to treat any mixed dataset as wholly nonpublic.

The economically meaningful distinction is whether the system helps competitors coordinate and police a common plan or helps each user make a better independent decision.

A Ban Built for the Biggest Landlords

Broad bans also impose long-term costs that disappear in a debate focused on this month’s rent.

Pricing is a major commercial use of machine learning because it generates frequent feedback. Developers can test forecasts, learn from errors, improve data systems, and build better tools for inventory, logistics, capacity planning, and demand estimation. Rules that exclude useful data or expose a provider to liability because two customers receive recommendations from similar models reduce the payoff from that experimentation.

The burden will fall unevenly. Large landlords can build proprietary systems using their own portfolios, engineers, and legal departments. Smaller owners are more likely to rely on third-party tools that spread development costs across many customers.

A ban on shared platforms may therefore protect firms large enough to develop the technology in-house while denying smaller rivals comparable capabilities. A rule sold as a check on large corporate landlords could end up strengthening them.

A state-by-state patchwork makes matters worse. Software providers build products for national markets. Faced with conflicting definitions of “nonpublic data,” “coordination,” and “algorithmic device,” they may design every product to satisfy the strictest state or stop serving smaller markets altogether.

Startups will struggle more than incumbents with those fixed compliance costs. The likely consequences include less entry, slower model improvement, and fewer experiments with tools that might lower costs or expand output.

That result also conflicts with the federal push for leadership in artificial intelligence. The White House’s “America’s AI Action Plan” identifies faster innovation and adoption as central to economic competitiveness. One need not endorse every part of that plan to see the tension. The federal government wants firms to develop and deploy artificial intelligence, while states prohibit major classes of learning and decision tools without requiring proof of competitive harm.

Antitrust errors do more than raise compliance costs today. False positives can redirect research spending and determine which technologies reach the market tomorrow.

Police the Cartel, Not the Code

The answer is not to ignore algorithmic cartels. It is to adopt a federal framework that targets the mechanisms of coordination while leaving room for beneficial experimentation.

First, antitrust law should keep agreement and competitive effects at the center of the analysis. An explicit arrangement among competitors to accept a common price, restrict output, or use a vendor to enforce discipline should face per se condemnation. A novel software system that generates recommendations while preserving independent decision-making should ordinarily receive rule-of-reason analysis. Courts should examine market power, adoption, actual operation, and efficiencies.

Parallel outcomes and common software may support an inference of collusion when paired with plus factors, meaning evidence that makes independent conduct less plausible. They should not replace proof of concerted action.

Second, the DOJ and FTC should issue guidance on algorithmic pricing. The agencies have already opened a 2026 inquiry into updated competitor-collaboration guidance and identified algorithmic pricing and information sharing as subjects for review. Any resulting guidelines should distinguish dangerous features from meaningful safeguards.

High-risk features would include a common commitment to follow recommendations, coercion or penalties for deviation, access to current and detailed rival-specific information, tools that expose deviations, communications among users about pricing strategy, uniform limits on discounts or price cuts, and broad adoption in a concentrated market.

The analysis should also consider the provider’s incentives, how often users receive and accept recommendations, whether outputs are individualized, and whether the system influences output or occupancy as well as price.

The agencies should pair that list with a rebuttable safe zone for systems that preserve independent rivalry. Relevant safeguards would include no user access to rival-specific data, aggregation and anonymization, appropriate delays, individualized objectives and outputs, meaningful freedom to reject recommendations, no penalties for doing so, no vendor-facilitated meetings among competitors, strong firewalls and audit logs, and documented efficiency justifications. Nonpublic data should count as one factor, not a forbidden category.

Third, federal enforcers should use statements of interest to clarify doctrine, rather than only to expand liability. The DOJ and FTC filed a 2024 statement of interest in hotel-pricing litigation explaining that an algorithm cannot immunize conduct that would otherwise violate antitrust law. That principle is sound.

Future filings should add its necessary limit. A common algorithm does not erase Section 1’s agreement requirement, and pooled nonpublic data do not by themselves establish a naked restraint. Carefully chosen filings could steer courts away from per se treatment of unfamiliar arrangements before economic evidence and experience support it.

Fourth, Congress should consider targeted preemption. A federal law could displace state and local rules that impose liability solely because firms use the same pricing algorithm, receive similar recommendations, or contribute nonpublic data without proof of agreement, coercion, or a likely anticompetitive mechanism.

States would remain free to enforce antitrust and consumer-protection laws against actual collusion, deception, discrimination, and unfair practices. The goal would be a coherent national baseline for interstate software and data services, not immunity for algorithms.

Preemption makes sense because fragmented state bans can impose costs beyond state borders. A prohibition in one large jurisdiction may dictate product design nationwide, discourage entry, and deny consumers elsewhere access to useful tools. Congress often adopts national rules when interstate commerce and networked technologies make 50 conflicting regimes unusually costly. Algorithmic pricing warrants the same consideration.

Finally, enforcement should remain empirical and open to revision. Agencies should study actual effects on prices, output, vacancies, quality, and entry before imposing permanent design rules. Remedies should address demonstrated sources of harm, such as rival-specific disclosures, coercive acceptance requirements, or monitoring tools, rather than banning entire categories of data.

Sunset provisions, regulatory sandboxes, and retrospective reviews would allow regulators to revise their approach as the technology and evidence develop.

Collusion Still Requires Collusion

New Jersey’s FAIR Act begins with a legitimate concern and ends with the wrong legal design. Algorithms can facilitate collusion. So can trade associations, consultants, phone calls, and private meetings. Antitrust law should target the agreement and the mechanisms that sustain it, rather than the general-purpose tool that carries the information.

The costs will extend beyond landlords losing a convenient product. Broad bans can make prices less informative, shield large incumbents from smaller technological rivals, encourage inefficient public disclosure, slow AI research, and fracture a national software market into incompatible state regimes.

Renters may pay through fewer available units, poorer matches, higher operating costs, and reduced investment. Meanwhile, zoning restrictions, construction costs, interest rates, and limited developable land will keep doing what they were doing before the algorithm arrived.

A sound policy would prosecute genuine hub-and-spoke cartels, provide clear federal guidance, examine actual conduct, create safe harbors for independent decision tools, and preempt state laws that mistake data processing for agreement.

Preserve rivalry. Demand evidence. And do not confuse better arithmetic with a cartel.

The post New Jersey’s War on Pricing Software Won’t Build More Apartments appeared first on Truth on the Market.

That is the logic running through two decisions the European Commission adopted last week involving the same company, under the same law, on the same day. Yet read side by side, they seem to come from different legal universes. 

The first measure, issued under Article 6(11) of the Digital Markets Act (DMA), requires Google to share anonymized search data with rival search engines and AI chatbots. The Commission surrounded that dataset with an elaborate system of safeguards. Identifiers must be removed. Access is delayed by at least a week. Rare or revealing queries are excluded. Eligible firms must pass screening, undergo an independent audit before receiving any data, and submit to annual audits thereafter. They must also use ring-fenced processing environments, comply with purpose and retention limits, and clear checks for sanctions and control by high-risk third countries. 

The second decision, issued under Article 6(7), requires Google to give rival AI assistants the same deep access to Android that Gemini receives. That includes ambient sensors, on-device app data, screen contents, and the ability to control other applications. The data concern identified users, include content, and arrive in real time. The main safeguard is a consent prompt. 

One dataset gets an armed escort. The other gets a checkbox. The DMA’s internal logic can explain the difference. A consequentialist analysis has a harder time doing so. 

A Walled Garden for Data, an Open Door for Devices

Understanding the mismatch requires a closer look at what each decision does. 

The search-data decision implements Article 6(11), which requires Google to provide rivals with ranking, query, click, and view data on fair, reasonable, and nondiscriminatory (FRAND) terms. The Commission begins with technical anonymization. Google must remove direct identifiers and timestamps, suppress rare terms and unusually long queries, and place each user in a group of at least 1,000 people who share the same location, device type, and language. In the Commission’s telling, the result is a “haystack” of disconnected queries. 

The decision goes much further. Because technical safeguards can only reduce the risk of reidentification, the Commission also regulates who may receive the data. Only genuine search businesses qualify. 

A firm must have operated in the European Union for at least two years. A newer entrant may qualify if it has raised more than €50 million and serves at least 50,000 monthly European users. The decision excludes sanctioned entities and firms controlled by third countries that pose structural cybersecurity or data-protection risks. Before sharing any data, Google may also assess whether a particular recipient presents serious cyber or privacy risks. 

Recipients must pass an independent audit before gaining access and submit to annual audits afterward. They must process the data in ring-fenced environments, use it only to improve search, refrain from training general-purpose AI models with it, and delete it on schedule. 

These safeguards largely turn on the recipient’s identity and conduct. The result resembles a walled garden, the same basic architecture gatekeepers built for themselves and that the Commission is dismantling elsewhere. Yet Article 6(11) itself devotes only one word to protection: “anonymized.” 

Article 6(7) takes a different approach. Its text expressly allows gatekeepers to adopt “strictly necessary and proportionate” measures to protect system integrity and security, provided they justify those measures. Yet in its March 2025 decisions involving Apple, the Commission reduced that protection to a narrow allowance. Interoperability solutions must work as well as Apple’s own and, according to the decision’s summary, “must not require more cumbersome system settings or additional user friction.” 

Under Article 6(7), any developer may request access, and Apple must process those requests within fixed deadlines. The framework does not allow firms to screen applicants based on identity, business model, or data-handling history. Draft joint guidelines from the European Commission and European Data Protection Board (EDPB) would go further by barring gatekeepers from considering an applicant’s record of violations under the General Data Protection Regulation (GDPR). 

The permitted safeguards focus on the product rather than the recipient. Gatekeepers may use consent prompts, encryption, and measures that preserve end-to-end encryption. They may not decide who gets through the door. 

Against that background, last week’s Android AI decision comes as little surprise. Google must support any integrity measure with “objective and verifiable evidence” of risk and apply it equally to its own services. As I argued in May, that standard makes precaution against genuinely novel threats almost impossible. 

The discrepancy is hard to miss. Where the statute says little about screening recipients, the Commission created an extensive vetting regime. Where the statute expressly permits protective measures, the Commission reduced them largely to prompts. 

To be sure, the provisions are worded differently, and that difference carries legal weight. Article 6(11) imposes an affirmative duty to anonymize the data, which the Commission had to translate into operational rules. Query logs also contain personal data, so the GDPR applies directly. Article 8(1) of the DMA requires compliance measures to respect those obligations. 

Article 6(7), by contrast, frames security protections as a limited exception available to the gatekeeper. The Commission therefore reads them narrowly to prevent evasion, as regulators often do with legal exceptions. The method is coherent enough. Commands receive broad effect. Exceptions receive little room to breathe. 

The Best Case for the Double Standard

The Commission’s different readings of the two provisions are defensible, at least up to a point. The strongest case rests on user control and scale. 

Start with user control. Under Article 6(7), each data flow begins with an identifiable user action, such as pairing a watch or tapping “allow.” That choice may provide a legal basis for the transfer through consent or contractual necessity. It also creates a transaction-specific safeguard through prompts, device-level permissions, and the ability to disconnect the service. 

Article 6(11) works differently. No individual user chooses to share anything. Query data from millions of people moves in bulk, and no workable consent mechanism could cover it. The only path to lawful disclosure is to remove the data from the GDPR’s reach through anonymization. 

Because anonymization carries so much legal weight, any threat to it becomes central. The most obvious risk is that a recipient could try to identify users again. Vetting, ring-fencing, and audits follow from that concern. The safeguards therefore track the legal basis for each form of access. Article 6(7) relies on the user’s choice. Article 6(11) relies on the recipient’s trustworthiness. 

Scale offers a second defense. Article 6(11) separates the amount of data disclosed from the recipient’s own size. Even the smallest eligible firm receives data derived from nearly the entire population of Google Search users, then keeps a copy in its own environment for up to five years. 

If anonymization fails, the harm can reach millions of users and cannot be undone. A leaked dataset cannot be unpaired, as AOL learned in 2006. Under Article 6(7), a third party receives data only from users who choose its service. Access grows with adoption, and users can revoke it. 

These arguments make a stronger case for the Commission’s distinction than the decisions themselves do. They still do not carry the day. 

When the Checkbox Becomes the Security System

That defense fails twice on its own terms, then a third time on a question it never asks. 

Start with consent. Nothing moves until a user pairs a device or taps “allow.” But Article 6(7) requires Apple and Google to open their interfaces and process requests before any user makes that choice. By the time the prompt appears, the requester already has technical access and may look no different to the user than a legitimate accessory maker. 

Attackers also work upstream of genuine consent. They can phish the prompt, spoof the accessory, or exploit the pairing process. As Miko?aj Barczentewicz warned before the DMA took effect, one determined bad actor can cause immense harm by exploiting an interoperability mandate. That risk weakens the case for treating privacy less carefully under Article 6(7) than under Article 6(11). 

The contrast becomes sharper when the regimes are placed side by side. Under the Google Search decision, data recipients must earn trust. Under the Apple and Android AI decisions, the Commission presumes it. Anyone may apply, and firms may not reject applicants based on who they are. The remaining check is a prompt that users have little basis to assess. 

We already know what an unvetted queue can contain. When Amazon reviewed applicants for DMA data access, more than 75% were based outside the European Union. Many appeared to be data brokers with murky privacy practices. Apple has reported interoperability requests broad enough to read every message and email on a user’s device. 

The regimes also differ in what changes hands. A sensible privacy framework should match safeguards to the sensitivity of the data. More revealing data should receive stronger protection. 

These decisions reverse that logic. Article 6(7) interoperability can expose the most sensitive material on a phone. The Android AI decision goes further by covering microphones, cameras, screen contents, and the ability of AI agents to imitate taps and complete transactions. Article 6(11), by contrast, covers week-old queries in which each individual is hidden among at least 1,000 others. The Commission has placed its lightest safeguards around the most sensitive data. 

Nor is it clear that Article 6(11) always creates greater risks at a larger scale. Imagine a popular smartwatch app with five million users that gets hacked or turns out to have been malicious all along. Five million people could have their messages exposed at once, tied to their identities. Article 6(7) requires no audit, certification, or background check to prevent that outcome. 

Now multiply that risk across hundreds of companies entitled to connect, each creating another point of failure. The unvetted crowd may pose a greater danger than the small number of heavily audited firms receiving Google’s search data. A regime that tracked actual risk would impose some checks everywhere and stronger ones where more people could suffer harm. 

Recipient screening is hardly a Brussels invention. Platforms used gated access, screened counterparties, and revocable permissions long before the DMA. They built those protections, presumably, to meet consumer demand for privacy and security. The Commission now overrides those design choices without adequately accounting for the functions they served. Those functions must then be rebuilt through regulation. 

Under Article 6(11), the Commission preserved that inherited machinery. Under Article 6(7), it dismantled it and replaced it with the one tool the rest of the statute treats with suspicion—a consent prompt. Elsewhere, the DMA assumes that user choice is weak and defaults are sticky. That is why it requires choice screens and restricts dark patterns. The same concern helps explain why markets often provide privacy and security through platformwide rules rather than one prompt at a time. 

The consequences are concrete. Siri AI will not launch in the European Union, and Gemini on Android could meet the same fate. When openness mandates and privacy law press in opposite directions, rational firms reduce functionality to limit their exposure. 

Fit for Purpose, Blind to Consequences

Why does Europe’s flagship digital law guard anonymous search queries like state secrets while leaving people’s messages, microphones, and screens behind a single tap? 

The two-speed approach makes sense only if consequences do not count. There is a legal explanation for the gap. The provisions use different language, and each decision follows its own internal logic. 

That may be defensible as statutory interpretation. It is nonetheless quite obviously poor privacy policy. The most sensitive data on a phone may receive the weakest protection, while less sensitive data comes wrapped in audits, screening, and access controls. No one designing a privacy regime from scratch would choose that result. Platforms had already built many of the missing safeguards under Article 6(7), and the Commission dismantled them. 

The stakes are rising. AI assistants can read screens, press buttons, access microphones, and spend money. Article 6(7) now helps determine who may wield those powers. The costs are already visible. Rather than open their most sensitive features to any requester, firms are withholding products from European users. 

The Commission’s own review called the DMA “fit for purpose.” That judgment depends on what the law actually produces. The search-data decision shows that the Commission knows how to protect users when it sees the risk. Its harder task is explaining why platforms may not do the same.

The post Europe’s Privacy Paradox: Fort Knox for Search Data, a Checkbox for Your Phone appeared first on Truth on the Market.

Sens. Amy Klobuchar (D-Minn.) and Eric Schmitt (R-Mo.) introduced the Securing Enforcement of Americans’ Right to Competition at Home (SEARCH) Act earlier this month. Billed as a bipartisan effort to protect consumers and curb Big Tech’s power over online search, the proposal would write into federal law many of the sweeping remedies sought in recent monopolization cases against Google.

The SEARCH Act would bar Google and other search engines from paying browsers and operating-system providers for default placement. It would require browsers and devices to present users with a “choice screen,” force search engines to provide rivals with data and search results at marginal cost, and restrict technology acquisitions and joint ventures more aggressively than current antitrust law does.

Protecting competition is a worthy goal. The bill’s economics are much shakier. By replacing the United States’ traditional focus on consumer welfare with European-style ex ante market design, the SEARCH Act would likely raise costs, weaken incentives to innovate, and give federal enforcers broad power to favor some firms over others.

Europe Already Tried This Search

The SEARCH Act borrows heavily from the European Union’s Digital Markets Act (DMA), which imposes broad, upfront restrictions on large digital platforms. Under the DMA, regulators need not prove that a company has market power or that a particular practice harms consumers. The law bans certain conduct categorically.

The DMA has now been in force for two years, and its promised gains in innovation and consumer choice remain elusive. In several cases, Europe’s ex ante rules have made familiar services more cumbersome.

After the European Union barred Google from integrating hotel and flight search tools in ways regulators considered “self-preferencing,” users had to click through more pages to complete basic bookings. Studies estimate that those extra steps cost European users millions of euros in lost time, while direct hotel-ad bookings fell 36%.

European rules also require “choice screens” that ask users to select a search engine or browser. These prompts may offer more options on paper, but many users ignore them, dismiss the pop-ups, and return to familiar services. More prompts do not necessarily produce more meaningful choice.

Regulatory uncertainty has also delayed new technology. European consumers waited longer for AI-powered search, live translation, and smart-assistant tools because platforms feared large fines under unclear DMA requirements. Those delays also left European technology startups without tools that their American competitors could already use.

Importing the same regulatory philosophy into U.S. antitrust law would make the American technology sector less competitive in the name of promoting competition.

There’s No Such Thing as a Free Default

Section 3 of the SEARCH Act would bar search engines from paying browsers, device makers, and other distributors for default placement. Yet default arrangements long predate the internet and often reflect ordinary, pro-competitive business conduct.

Supermarkets, for example, routinely give their private-label products prime shelf space. That arrangement can reduce “double marginalization,” which occurs when firms at different stages of a supply chain each add their own markup. Fewer markups can mean lower prices for shoppers.

In digital markets, payments for default status can serve as a form of “competition for the market.” Search engines compete for distribution by offering browsers and device makers better terms. The resulting revenue helps fund free software, including open-source browsers such as Mozilla Firefox, as well as mobile operating systems and privacy-focused services.

A ban would strip independent platforms of a major source of revenue. They would then have to cut features, add more advertising, charge users, or some combination of the three. Consumers would pay the price through higher costs, weaker products, and a worse browsing experience.

A Search Engine on Someone Else’s Dime

The SEARCH Act’s data-sharing rules are just as troubling. Sections 5 and 6 would require covered search platforms to provide “qualified competitors” with access to search indexes, ranking signals, and user-query data at marginal cost.

Building a modern search engine takes billions of dollars and years of continuous work. Companies must crawl the web, organize vast amounts of information, and refine the algorithms that decide which results users see.

Requiring firms to share those assets at cost creates a free-rider problem. Rivals gain access to expensive infrastructure without making comparable investments, which gives them less reason to develop better crawling tools or new search designs. Incumbent search engines also have less reason to improve their products when regulators can require them to hand those improvements to competitors.

Mandatory data sharing also creates serious privacy risks. Even when firms remove obvious identifiers, distributing search-query data across more companies gives hackers more targets and creates more opportunities to identify users by combining datasets.

The SEARCH Act would trade user privacy and long-term investment for a government-designed shortcut to competition.

Regulating the Search War That AI Already Changed

The SEARCH Act arrives just as generative artificial intelligence is reshaping online search more quickly than any law or lawsuit could. The rapid growth of conversational tools such as ChatGPT, Claude, and Perplexity shows how fast technology markets can change when new products challenge established ones.

These services have already brought real, market-driven competition to traditional search engines without Congress designing the terms of entry.

The SEARCH Act would still pull “AI search” into its regulatory scheme. Section 3(e) would require covered platforms to obtain written approval from the Federal Trade Commission before making certain acquisitions, investments, or joint ventures involving search or AI competitors.

That preapproval requirement would add delay and uncertainty to venture funding, strategic partnerships, and product development. It could slow the deployment of American AI models and weaken U.S. technological leadership just as competition from China and other geopolitical rivals intensifies.

Why Courts Need Room to Judge

Flexible U.S. antitrust law has one major advantage over rigid ex ante mandates. Judges can weigh competitive harms against benefits, account for commercial and technological realities, and tailor remedies to the conduct proved at trial. They also know that an appeals court can reverse them if they misapply the law or give too much weight to one side of the evidence.

Consider the recent Google Search case. U.S. District Judge Amit Mehta found that Google’s default agreements with Firefox and Safari helped it unlawfully maintain its search monopoly. Those contracts limited rival search engines’ ability to reach “minimum efficient scale,” the level of activity needed to compete economically.

According to Mehta, Google’s default placement generated more searches, clicks, and user data, which helped the company refine its index and ranking algorithms. Rivals received less data and therefore struggled to produce comparably strong search results.

When fashioning a remedy, though, Mehta confronted tradeoffs that the SEARCH Act largely ignores. He recognized that simply banning Google’s distribution agreements could harm competition among both search engines and browsers. He also concluded that requiring Google to provide its data to rivals for free or at marginal cost could weaken incentives to create and improve those assets, a concern long reflected in U.S. antitrust law.

Mehta instead required Google to offer certain data to competitors on “commercial terms.” That approach restricts Google’s control over its proprietary resources while still requiring rivals to pay market-based compensation. It seeks to preserve incentives for Google and its competitors to collect data, improve search quality, and develop competing tools.

The EU’s DMA offers no comparable flexibility. Neither would the SEARCH Act’s categorical ban on default agreements.

Google has appealed the liability ruling, which legal scholars have also criticized. Appellate review gives courts another chance to test whether a practice’s anticompetitive harms truly outweigh its pro-competitive benefits. Rigid ex ante rules sharply limit, or eliminate, that safeguard.

Europe’s Mistakes Need No U.S. Sequel

America’s technology sector became a global engine of growth under antitrust laws that focus on consumer welfare, rather than firm size or the protection of competitors from vigorous competition.

Existing laws, including the Sherman Antitrust Act, already give federal courts power to police anticompetitive conduct case by case. Recent platform litigation shows that judges can examine evidence, weigh competitive harms against benefits, and craft targeted remedies without imposing broad rules that may leave consumers worse off.

Congress should learn from Europe’s experience before importing its mistakes. The United States does not need antitrust rules that assume the verdict before the evidence arrives. It needs enforcement that follows the facts wherever they lead.

The post The SEARCH Act Copies Europe’s Homework—and Its Mistakes appeared first on Truth on the Market.

Competition authorities often warn that procedure should not become an obstacle to enforcement. Fair enough. Antitrust investigations can be slow, document-heavy, and vulnerable to delay tactics. A firm with deep pockets may try to turn “process” into attrition.

But that is only half the institutional problem. An agency with broad discretion, weak disclosure duties, and limited independent review can turn enforcement into regulation by accusation.

That concern is economic as well as legal.

Due process helps competition authorities separate harmful conduct from vigorous competition. Notice, access to evidence, a meaningful chance to respond, protection for privileged and confidential information, reasoned decisions, and independent review all improve the quality of enforcement. When those safeguards are weak, agencies make more errors, firms face greater uncertainty, and interest groups gain more opportunities to seek favorable treatment.

The costs do not stay inside the hearing room. They appear in lower investment, less experimentation, and fewer innovations.

My recent Antitrust Chronicle article on the “next step forward” examined the substantial international framework already in place. It includes the Organisation for Economic Co-operation and Development’s (OECD) recommendation on procedural fairness, the International Competition Network’s (ICN) Framework for Competition Agency Procedures, and the growing use of due-process commitments in trade agreements.

The problem is no longer a shortage of principles. It is the gap between stated commitments and actual practice.

That gap reflects institutional incentives. Competition agencies do not operate as detached maximizers of social welfare. Like other institutions, they respond to political pressure, resource limits, internal priorities, and organizational self-interest.

A serious reform program must therefore do more than urge agencies to be fair. It must make sound procedure visible, reward compliance, and impose some cost when authorities ignore it.

Due Process Is an Error-Cost Problem

The law & economics case for due process begins with uncertainty. Outside the easy case of a naked cartel, the same conduct may help or harm competition depending on facts that are difficult to observe.

Exclusive contracts may shut out rivals, or they may encourage relationship-specific investment. Product integration may exclude complementary products, or it may lower transaction costs and improve quality. A merger may reduce rivalry, or it may combine the assets needed to bring a new product to market.

The error-cost framework emphasized by scholars at the International Center for Law & Economics (ICLE) asks decisionmakers to minimize the combined social costs of false condemnation, false acquittal, and administration. Scholars usually apply that framework to substantive rules, burdens of proof, and legal presumptions. It applies just as readily to procedure.

Procedural safeguards improve the information available to the decisionmaker. Early, specific notice tells the investigated party which theory it must answer. Access to inculpatory and exculpatory evidence reduces the risk that an agency will build its case on a partial record. The right to counsel and adequate response time allow technical and economic claims to be tested rather than merely asserted.

A reasoned decision requires the authority to connect the evidence to the governing legal standard. Independent review adds an external check on confirmation bias and the gradual expansion of an agency’s mission.

These protections cannot guarantee correct outcomes. They can, however, reduce both the likelihood and the cost of error. That is especially valuable in innovation-driven markets, where competitive effects are hard to predict and a mistaken remedy may shape business decisions for years.

In “Innovation and the Limits of Antitrust,” Geoffrey Manne and Joshua Wright explain why authorities may condemn novel business practices before their benefits become clear. Sound procedure can slow that rush to judgment while preserving enforcement against genuine exclusion.

The economic value of procedure also extends beyond individual cases. Predictable enforcement lowers the risk associated with investment and allows firms to plan transactions, contracts, and product launches with greater confidence. It also encourages voluntary compliance because firms can better understand what the law requires.

Fair procedures can also strengthen an agency’s legitimacy. Firms are more likely to accept adverse decisions when they can see how the agency reached them, rather than dismissing them as political acts.

Christopher Yoo, Thomas Fetzer, Shan Jiang, and Yong Huang make the broader case in their comparative study, “Due Process in Antitrust Enforcement.” They connect procedural protections to accuracy, government credibility, economic growth, limits on bureaucratic abuse, competitive reform, and lower corruption.

Those benefits are central to the institutions that support a market economy.

Due Process, Terms and Conditions Apply

Internationally, the normative work is largely complete. The OECD’s Recommendation on Transparency and Procedural Fairness in Competition Law Enforcement calls for predictable rules, impartial institutions, nondiscrimination, proportionality, timely resolution, meaningful engagement with parties, protection of privileged and confidential information, independent review, and periodic reassessment of enforcement practices.

The ICN’s Framework for Competition Agency Procedures turns similar commitments into an agency-led, voluntary framework. Participating authorities publish templates describing their procedures, agree to core principles, and may take part in interagency dialogues. The current CAP 2.0 work plan aims to reaffirm those commitments, draw attention to neglected areas, and recruit more signatories.

That is genuine progress. Soft law can shape expectations, help reform-minded officials press for change within their own governments, and create a shared vocabulary across legal systems. The CAP templates also make national procedures easier to compare.

Yet soft law has familiar limits. The framework creates few enforceable rights for the firms and individuals directly affected by agency action. Interagency dialogues are generally confidential and depend on officials’ willingness to participate. Authorities may invoke national law to justify departures from the framework. Periodic reviews tend to celebrate collective progress more readily than identify specific failures.

As James Rill and Jana Seidl argue in their analysis of the CAP’s “promise of accountability,” genuine convergence requires a mechanism that reveals whether jurisdictions have actually implemented the norms they accepted.

Other contributions to the CPI Antitrust Chronicle’s June 2026 issue on due process make the same point. J. Mark Gidley and Daniel Sokol argue that formal rights mean little if parties cannot invoke them when they matter most, including during searches, evidence collection, formulation of charges, interim relief, and review.

Ian Forrester and Pablo Trevisan compare administrative and adversarial systems while stressing the practical value of being heard by a neutral decisionmaker. John Taladay and Christine Ryu-Naya warn that market investigations can expand beyond fact-finding and become vehicles for industrywide remedies without procedural protections equal to that power. Abbott (Tad) Lipsky points to the painfully slow movement from decades of discussion to measurable improvement.

The recurring defects are well known. Agencies may rely on vague or shifting theories of harm, restrict access to the file, withhold exculpatory evidence, impose compressed response periods, or conduct intrusive searches with weak ex ante or ex post controls. They may use confidentiality rules to obscure their reasoning, allow investigations to drag on until the process itself becomes punishment, impose interim measures before the facts are settled, or face judicial review too deferential, slow, or narrow to offer a useful remedy.

A consensus on paper does not produce compliance on its own. Explaining the gap requires asking a less comfortable question. Who benefits from weak process, and who pays for it?

Why Agencies Prefer Procedural Wiggle Room

The first obstacle is a collective-action problem. Better procedure produces broad benefits, including greater legal certainty, fewer errors, more investment, and stronger public confidence. Consumers, businesses, and the wider economy share those gains. Enforcement agencies and political institutions bear the immediate costs.

Disclosure consumes staff time. Hearings slow decisions. Independent review can overturn agency victories. Clear standards constrain discretion. An agency may therefore see procedural reform as surrendering authority in exchange for benefits that accrue mostly to others. Six factors help explain the resistance.

First, public-choice analysis predicts this tension. My recent Truth on the Market essay on why administrable antitrust presumptions can become “too useful to kill” explains that agencies naturally prefer rules that reduce evidentiary burdens, conserve resources, and improve win rates. The same logic applies to procedure.

An authority gains an institutional advantage when it can formulate allegations late, restrict access to the file, negotiate remedies under severe time pressure, or combine investigation with the initial decision. It has little reason to surrender those advantages voluntarily.

Second, weak process creates opportunities for rent seeking. Less-efficient rivals, domestic incumbents, organized labor, sector regulators, and political officials may all benefit when an agency investigates a successful firm, even if the competition case is thin. Competition law can become a tool for redistributing rents, pursuing industrial policy, or burdening foreign businesses.

The targeted firm bears concentrated costs. Political beneficiaries can claim to defend “fairness,” small businesses, jobs, or national sovereignty. Broad substantive standards and opaque procedures make consumer protection difficult to distinguish from competitor protection.

Third, agency-led international institutions inherit their members’ incentives. The ICN’s strength comes from its practical expertise and freedom from the slow machinery of treaty negotiations. Its weakness follows from the same design. An organization composed of enforcement agencies is unlikely to build a demanding system for publicly censuring those agencies.

Confidential peer dialogue and consensus-based recommendations are the politically feasible price of participation. They also help explain why the network’s implementation mechanisms remain modest.

Fourth, institutional capacity varies widely. A mature agency may be able to employ hearing officers, separate investigative and decisional staff, issue detailed decisions, and support expedited judicial review. A newer authority with a small budget may struggle to translate documents, preserve digital evidence, and meet basic deadlines.

Requiring every authority to adopt an elaborate procedural model could divert scarce resources from cartel enforcement and competition advocacy. Yet limited resources can also become a permanent excuse for indefinite investigations and minimal disclosure.

Fifth, due process becomes most vulnerable when political pressure makes it most necessary. High-profile cases against large or unpopular companies create demands for speed and visible action. Officials who defend procedural protections may be accused of serving monopolists or foreign interests.

The costs of a mistaken intervention often emerge slowly and are difficult to trace back to the agency. The political rewards of an aggressive announcement arrive immediately. That mismatch encourages agencies to act first and learn later.

Sixth, legal tradition matters. Some systems place considerable trust in integrated administrative decisionmaking. Others rely more heavily on adversarial courts. Critics may therefore portray procedural convergence as an effort to export American litigation practices.

That objection has force when reformers demand institutional uniformity. It weakens when the goal is functional equivalence. Whatever the institutional design, parties should receive timely notice, meaningful access to the case against them, an opportunity to respond, protection against conflicts of interest, and effective independent review.

Putting Teeth in Procedural Fairness

A workable reform agenda should begin with modesty. Due process cannot become a license for endless delay, tactical discovery, or obstruction of legitimate investigations. National systems also need not copy U.S. federal litigation. The aim is narrower—to improve the accuracy and legitimacy of decisions at a reasonable administrative cost.

First, make performance visible. The Framework for Competition Agency Procedures templates should evolve beyond descriptive questionnaires and include a limited set of comparable indicators. Agencies could report median investigation length, when they disclose theories of harm, rules governing access to inculpatory and exculpatory evidence, the use and duration of interim measures, access to privilege review, publication of reasoned decisions, and the time required to obtain judicial review.

These indicators need not produce a crude league table. They should reveal where formal commitments diverge from actual practice.

The indicators need not produce a simplistic league table. Their purpose is to reveal where formal commitments diverge from operational reality.

CAP reviews could also publish anonymized case studies and aggregate statistics on interagency dialogues. A process that identifies persistent patterns without turning every dispute into diplomatic theater would raise the reputational cost of noncompliance. Reform-minded agency leaders would also gain an external benchmark when seeking legislative authority or additional resources.

Second, give affected parties a structured voice without handing them control of the framework. Firms should not direct peer review, but they should be able to submit documented procedural concerns to an independent CAP contact point or the OECD secretariat.

Bar associations, economists, consumer groups, and nongovernmental advisers could help screen complaints and identify recurring problems. Requiring specific allegations and exhaustion of domestic remedies would discourage tactical filings. The process could produce a confidential referral, an anonymized thematic review, or a public recommendation when the evidence shows a persistent pattern.

Third, strengthen internal checks that cost less than a full institutional redesign. Agencies can appoint hearing officers or procedural ombudsmen to resolve disputes over access, privilege, deadlines, and confidentiality. They can require written approval and prompt review for interim measures and dawn raids—that is, unannounced searches of business premises.

Agencies can also separate investigative staff from initial decision makers where practical, adopt presumptive deadlines with public explanations for extensions, and require decisions to address material defense evidence rather than merely repeat the agency’s theory.

These reforms can fit agencies of different sizes. A small authority need not create a separate court. It could rely on a rotating external panel, a ministry-wide administrative judge, or a regional cooperation arrangement. Technical assistance should support these institutional safeguards alongside training on how to bring cases.

Fourth, attach limited procedural conditions to international cooperation. Evidence sharing, confidentiality waivers, and investigative assistance expand enforcement power. Those tools should come with assurances that shared information will receive adequate protections for privilege, confidentiality, permitted use, and review.

Countries need not suspend cooperation whenever procedures differ. But particularly sensitive assistance should depend on verified safeguards and should be restricted when the receiving authority cannot provide them.

Fifth, use trade agreements with greater care and credibility. Modern agreements already offer models for a procedural baseline. Chapter 21 of the United States-Mexico-Canada Agreement includes commitments on notice, representation by counsel, the opportunity to present evidence, written decisions, and review. OECD research documents the growing use of competition provisions in trade agreements. The African Continental Free Trade Area Competition Protocol points toward a more integrated model that includes state-to-state dispute settlement.

Future agreements should make a narrow set of procedural commitments subject to consultation and, in cases of systematic violation, dispute settlement. Remedies should be targeted and graduated. Transparency requirements, corrective-action plans, suspension of discretionary cooperation, and prospective compliance commitments will usually serve the goal better than tariffs. Tariffs burden consumers and can turn a rule-of-law dispute into a protectionist contest.

Section 301 of the Trade Act of 1974 and similar tools may provide pressure in extreme cases. They should remain a last resort.

Sixth, improve judicial review when timing matters most. Review that arrives after a transaction has collapsed or a technology has been redesigned offers little practical protection. Courts should be able to examine procedural irregularities, test the evidentiary basis for interim measures, and grant expedited relief when irreversible harm is likely.

Deference to agency expertise should not extend to undisclosed evidence, shifting legal theories, or remedies disconnected from the agency’s findings.

Seventh, evaluate procedure as economic policy. Agencies routinely assess the effects of business conduct. They should also assess the effects of their own processes.

Ex post reviews could examine whether preliminary theories survived, how often remedies were modified or reversed, how long investigations lasted, and whether compliance costs matched the stakes. The aim is to encourage institutional learning and identify patterns that ordinary case-by-case review may miss, rather than punish agencies for reasonable losses.

When Bad Process Becomes Bad Economics

The international consensus on procedural fairness is broad. The incentives to implement it remain weak.

Arguments for better process often focus on dignity, legality, and fairness. Those values matter. So do the economic consequences of getting procedure wrong.

In dynamic markets, enforcement can redirect innovation before courts have a chance to review it. Interim restrictions, design mandates, forced access, limits on integration, and merger delays can alter investment decisions long before the evidence is complete.

Toshiaki Takigawa and Dirk Auer recently made this point in discussing artificial intelligence regulation. Preventive intervention imposes the costs of error early. The faster technology changes, the less useful a remedy may become by the time the record is complete, and the more damage a mistaken early order may cause.

Global enforcement compounds the risk. A multinational firm may adopt the rules of the most restrictive jurisdiction worldwide because maintaining separate products or business models is costly. A procedurally defective decision in one country can therefore reshape conduct far beyond its borders.

Smaller firms and startups face even greater pressure. They have fewer resources to endure lengthy investigations, contest confidentiality claims, or pursue appeals in several jurisdictions. Legal uncertainty can discourage the entry and experimentation that competition law is supposed to protect.

Due process should therefore be treated as neither a concession to defendants nor a box-checking exercise for international organizations. It improves the information available to decisionmakers, constrains discretion, and reduces the social cost of enforcement errors.

The OECD and the ICN have built a valuable consensus. The next step is to align incentives with that consensus through measurement, independent feedback, conditional cooperation, targeted trade commitments, and effective review.

No reform will eliminate political pressure, institutional self-interest, or differences among legal systems. A realistic agenda can still make procedural failure more visible, make sound practices easier to adopt, and attach credible consequences to persistent noncompliance.

Competition law protects rivalry and discovery. Its own procedures should show the same humility about what authorities know and the same caution about shutting down experimentation.

When process becomes an afterthought, competition can become collateral damage.

The post Competition Law’s Fairness Gap: Why Better Procedure Requires Better Incentives appeared first on Truth on the Market.

The European Commission’s latest Digital Markets Act (DMA) decisions capture that tension. Europe wants more innovation, investment, and globally competitive digital platforms. Yet when a company assembles the data, technology, distribution, and complementary services needed to compete worldwide, Brussels increasingly treats those advantages as inputs to be shared with rivals.

On July 16, the Commission adopted binding measures requiring Google to give eligible search engines, including AI chatbots with search functions, access to anonymized Google Search data. A parallel decision requires Google to open 11 categories of Android functionality to competing AI assistants. These include sensor inputs, app data, operating-system controls, screen automation, on-device computing resources, background execution, and hotword activation.

The Commission says these mandates will promote contestability, consumer choice, and innovation. The immediate gains are easy to see. Rivals receive valuable inputs, developers gain access to Android features, and consumers may encounter more services competing within Google’s platform.

The costs are harder to spot and easier to ignore. Forced access may weaken incentives to build proprietary datasets, reduce returns on risky platform investments, increase cybersecurity risks, and push competing systems toward the same design. It may also lead companies to delay, degrade, or withhold future products in Europe.

These decisions therefore deserve scrutiny as more than access mandates. They alter the rules governing who bears the cost of innovation and who receives the reward.

The Algorithm by Other Means

The Commission stresses that Article 6(11) does not require Google to disclose its source code, algorithms, or technology. Formally, that is true. The mandate instead covers anonymized ranking, query, click, and view data generated through paid and unpaid search.

Google must generally provide information it collects and uses to improve Search, including entered queries, language and device metadata, viewed URLs, user interactions, and result positions. The public measures contemplate a delay of at least seven days and allow each eligible recipient to receive the data for up to five years.

That source code remains private does not resolve the intellectual-property problem. In a data-driven business, valuable know-how appears in more than written algorithms. It also emerges from the feedback generated when those algorithms operate across billions of searches.

Query reformulations, click patterns, result positions, dwell times, and other behavioral signals can reveal how a system reacts to particular inputs. Combined with a recipient’s own models, experiments, and datasets, that information can reduce the cost of approximating aspects of the incumbent’s ranking behavior.

The meaningful distinction concerns direct access to technical instructions and access to the behavioral data from which competitors may infer parts of those instructions. The second form of access will not recreate Google Search. Search quality also depends on crawling, indexing, engineering talent, experimentation, brand, distribution, computing infrastructure, and other complementary assets.

The International Center for Law & Economics (ICLE) has therefore cautioned that shared data may not transfer much value and that the returns from additional search data may decline. Even so, mandated access can capture part of the return on Google’s past investment by lowering rivals’ costs of experimentation, error correction, and imitation.

Search data did not materialize on its own. Google generated it through decades of investment in crawling, indexing, infrastructure, cybersecurity, user interfaces, quality testing, and fraud prevention. Individual facts may not qualify as intellectual property, but the organized stream of information produced by the platform remains an economically valuable asset. Its value reflects both Google’s technological investment and the trust of users who chose to submit their queries.

The Commission’s pricing rule deepens the problem. Google generally may recover only the incremental costs of preparing, storing, and transmitting the dataset, plus a return on incremental capital capped at its weighted average cost of capital. An additional margin is available only in exceptional circumstances.

That formula largely excludes the sunk costs of creating the underlying information asset and the option value of keeping it exclusive. For information goods, those omissions are substantial. Innovation-intensive products often require high fixed and sunk costs, face uncertain commercial prospects, and cost little to reproduce once created.

A price based mainly on delivery expenses therefore ignores the investments that created the value being shared. As ICLE has warned, a fair, reasonable, and nondiscriminatory (FRAND) access rule can become a subsidy for competitors.

The decision does not amount to uncompensated expropriation. Access remains limited, Google must anonymize the data, recipients must meet eligibility requirements, and some compensation is available. Economically, though, the mandate resembles a compulsory license priced without full regard for the investment that produced the licensed advantage.

It transfers part of Google’s informational returns to rivals while leaving Google responsible for the costs, liabilities, and security risks of creating and maintaining the asset.

The Mandate That Keeps on Mandating

The Commission argues in its search-data Q&A that Google will retain ample incentives to innovate. Recipients will receive only a modified subset of its data, may not systematically reproduce Google’s search results, and may not use the dataset to train general-purpose AI models. Those safeguards limit the mandate’s reach, but they do not establish that it will leave investment unchanged.

Investment incentives operate at the margin. A rule need not capture every return to alter a company’s behavior. It need only reduce the expected payoff from the next dollar spent improving search quality, collecting data, developing behavioral signals, or launching a new feature.

A company that expects regulators to make valuable feedback available to competitors at a controlled price may invest less in producing that feedback. It may also reserve its most commercially sensitive innovations for products or jurisdictions where it can retain more of the resulting value.

The effects reach beyond Google. An aspiring platform must now account for the possibility that success in Europe will turn the scale economies it created into duties to assist rivals. Investors will discount expected returns to reflect that risk. Some firms will still invest, but they will demand higher projected returns before doing so. Marginal projects will lose funding, ambitious forms of integration will become less attractive, and business models that depend on long-term cross-subsidies will become harder to finance.

The mandate also creates a predictable regulatory ratchet. If recipients fail to gain market share, they can argue that the dataset is too narrow, delayed, anonymized, or expensive. If privacy protections make rare queries less useful, beneficiaries can press the Commission to relax them. If access does not allow rivals to match Google’s quality, they can demand more data fields, shorter delays, or technical assistance.

ICLE warned in its comments on Article 6(11) that judging “effective compliance” by competitors’ commercial results rather than the availability of workable access would encourage exactly this progression. Each disappointing competitive outcome could become evidence that the previous mandate did not go far enough.

The Commission’s own Q&A gives that concern some force. It rejected Google’s initial implementation in part because the company removed 90% to 100% of unique queries and attracted little meaningful uptake. Uptake may help show whether rivals find the data useful, but it does not measure consumer welfare. A dataset can satisfy an access requirement even when competitors decide it lacks commercial value. A highly valuable dataset can also impose greater costs on privacy, security, and future innovation.

Once competitor use and market share become measures of regulatory success, the Commission moves beyond opening access. It begins designing the competitive outcome.

Anonymized, With an Asterisk

The Commission has taken privacy seriously. Its system removes direct identifiers and some metadata, suppresses rare or unusually long queries, generalizes locations, aggregates interaction times, and strips out advertising URLs. Contracts, independent audits, limits on onward transfers, retention rules, and use restrictions add further protection.

Those safeguards matter. The decision does not authorize an unrestricted transfer of identifiable search histories. It also does not automatically open the data to every Chinese, Russian, or other foreign-controlled company. The Commission allows Google to exclude sanctioned entities and firms controlled by countries that pose serious and structural cybersecurity or data-protection risks. It also permits public-security exemptions and requires processing in the European Economic Area or under protections deemed essentially equivalent.

These measures reduce the danger. They do not make it disappear.

Search queries can reveal sensitive information about health, finances, travel, employment, political activity, intimate relationships, business strategy, and personal vulnerabilities. Removing direct identifiers may not prevent recipients from combining query data with their own logs, public records, advertising data, leaked datasets, or inference models.

Mikolaj Barczentewicz’s analysis of the preliminary proposal made the central point. Anonymity depends on more than what appears in the dataset itself. It also depends on the auxiliary information and technical capabilities available to each recipient.

That concern follows directly from Article 6(11)’s competitive premise. Recipients are expected to possess the complementary tools needed to extract commercial value from the data. The same expertise may also help them reconstruct sensitive information. Contracts and annual audits can deter abuse, but they may detect it only after data have been copied, inferred, compromised, or transferred.

National-security concerns also require precision. The public materials do not show that raw, identified searches by U.S. military officials or diplomats will reach hostile governments. But searches by American personnel in Europe may still form part of the larger pool from which the shared dataset is drawn. If the data remain sufficiently granular or are poorly anonymized, they could reveal patterns involving travel, facilities, vendors, technical problems, operational interests, or personnel concerns.

The Commission’s high-risk-country and public-security provisions address some recipient risks, but they do not necessarily match U.S. national-security judgments. They also cannot eliminate cyber intrusions, hidden beneficial ownership, acquisitions after certification, insider threats, or inferences drawn from multiple lawful datasets.

Google’s July 16 response argues that the measures expose private searches without adequate anonymization, knowledge, or consent and create risks to privacy, business secrets, and national security. That remains Google’s position rather than an adjudicated finding. Even so, the costs it identifies deserve analysis rather than dismissal as the grumbling of a regulated firm.

Privacy and security are part of service quality. A competition rule that expands rival access by weakening users’ expected confidentiality may produce more competitors and a worse product.

A Remedy With Guardrails—and One Without

The American search-remedies litigation offers a useful comparison. The U.S. district court imposed its remedy in United States v. Google only after finding specific exclusionary conduct. Its user-side data mandate covers the underlying information used the Generalized Learned User Embeddings (GLUE) and RankEmbed models and requires at least two disclosures. The court will determine the final number and timing after consulting a technical committee.

The judgment also draws firm boundaries around what Google must share. It excludes algorithms, ranking signals, scores, post-trained large language models, intellectual property, and trade secrets.

Recipients must satisfy court-approved security standards, pass regular audits, show a credible plan to invest and compete, undergo annual recertification, and pose no threat to U.S. national security. Google may object to a recipient’s initial or continued eligibility and receive a hearing. The court also recognized that anonymized user-side data could remain highly sensitive, so it restricted sales and onward sharing.

The American remedy still invites criticism. It prices access at marginal cost and may weaken incentives to invest. Yet it remains more closely tied to adjudicated conduct, defined datasets, judicial oversight, trade-secret protection, national-security screening, and recipient-by-recipient review.

The DMA decision reaches further in both duration and ambition. It does not respond to a finding that Google unlawfully acquired a particular dataset. It treats Google’s continuing accumulation of search data as a structural advantage that regulators may redistribute for as long as Search remains designated.

The decision also extends eligibility to AI chatbots with search functions and seeks, where technically feasible, to approximate the methods and speed with which Google uses the data itself. The U.S. judgment remains an antitrust remedy, even if an interventionist one. The European model looks increasingly like continuing public-utility regulation of information.

Leveling the Platform Down

The Android decision presents a related but distinct problem for innovation. The Commission requires Google to give third-party AI services free access, on equally effective terms, to Android functions involving invocation, context, actions, and computing resources.

That access is extensive. Third-party assistants may receive real-time input from microphones, cameras, screens, and speakers. They may interact with apps and Google services, automate tasks in virtual windows, change system settings, use on-device AI models, run in the background, and activate through always-on hotwords. Google must document the interfaces, provide testing and technical support, extend future functionality to third parties, and avoid unnecessary friction for users.

These rules could help independent AI developers. An assistant with deep device access can do far more than one confined to an ordinary app sandbox, the restricted environment that limits what an app may see and control. Consumers may gain more choice, and developers may create features Google would never have pursued.

The longer-term cost is weaker competition between integrated systems. Tight integration between Gemini and Android is itself a product feature. It gives Google a way to challenge rival AI providers and distinguish Android from Apple’s model. Requiring Google to reproduce those advantages for competitors may increase competition within Android while reducing competition between rival AI systems.

ICLE’s comments on Article 6(7) identify this neglected distinction. Giving equivalent Android access to leading AI providers could soften competition across the broader AI market by limiting Google’s ability to use operating-system integration as a competitive response. The main beneficiaries may be powerful incumbents rather than scrappy startups. They receive a regulated route into a platform they did not build.

Security risks make equal treatment harder still. Google can inspect its own code, impose internal development rules, revoke credentials immediately, test security throughout the development process, and coordinate responses across Android. It cannot exercise the same control over every outside provider. Equal technical access does not produce equal risk.

The Commission permits integrity measures that are strictly necessary and proportionate. For several sensitive functions, it also allows objective eligibility requirements and independent certification. Yet rules demanding transparent, objectively verifiable, and broadly symmetric restrictions may fit poorly with new threats.

Security teams often must act before they possess conclusive proof. Waiting for demonstrated exploitation can mean waiting until users have already suffered harm. ICLE therefore warns that Google may face a stark choice. It can expose sensitive functions broadly or remove them from its own services so that equal-treatment duties no longer apply.

That creates a predictable form of defensive leveling down. Google may narrow functionality, reduce the sensitivity of available application programming interfaces, delay integrated features, or withhold some services from Europe. Smaller platform developers watching the result may avoid close integration altogether.

When Brussels Becomes the Product Manager

My July 8 analysis of the Court of Justice of the European Union’s (CJEU) Google Android judgment offers a broader frame for evaluating these decisions. Intervention becomes especially risky when regulators treat successful platform design as presumptively suspect and fail to ask what would have existed without the challenged integration.

Android’s compatibility rules, monetization arrangements, defaults, and complementary services helped coordinate handset manufacturers, developers, advertisers, and users. Weakening one part can impose costs on the others.

The July 16 decisions extend that logic beyond retrospective antitrust liability. They prescribe data fields, anonymization methods, recipient eligibility, prices, interface design, access quality, documentation, technical support, timelines, and future functionality. This goes well beyond barring exclusionary conduct. It amounts to continuing administrative control over platform technology.

My related analysis of the DMA’s rule-of-law problems describes this shift from competition law toward product management and industrial administration. Because the DMA does not require the same case-specific proof of competitive harm as traditional antitrust law, procedural discipline and economic evidence become more important.

The formal specification proceedings ran from Jan. 27 to July 16, with preliminary search measures released April 16. The Commission says the process followed two years of discussions with Alphabet and included consultations, testing, and expert input. Google therefore received a process, but the deeper concern remains.

A compressed administrative proceeding under a statute designed to bypass much of traditional antitrust’s effects analysis now determines technical, privacy, security, and intellectual-property questions whose consequences may extend far beyond one product cycle.

The asymmetry in error costs should give regulators pause. The state can revise a mistaken rule. Google cannot fully reverse a data disclosure, undo a security breach, or recover innovation returns once regulators have transferred them to competitors. Under those conditions, “move fast” makes for a poor governing principle.

A Mandate With Brakes

A more economically defensible approach would treat dynamic competition, privacy, security, and firms’ ability to retain returns on investment as core design constraints.

Search-data access should begin narrowly. The Commission should tie it to evidence of specific competitive bottlenecks and initially limit disclosure to synthetic, sampled, or heavily filtered data. It should cap the frequency and duration of access, then expand the mandate only when evidence shows consumer benefits. Competitor uptake alone is a poor measure.

The ban on systematic replication also needs teeth. Recipients should face enforceable limits on reverse engineering, model extraction, combining the data with outside datasets, onward transfers, and changes in ownership or control.

Eligibility should include a recipient-specific national-security review conducted with relevant allied authorities. General country-risk categories cannot capture hidden ownership, personnel access, cybersecurity capabilities, prior incidents, acquisition risk, or the sensitivity of a proposed use. Google should also have a meaningful chance to challenge access before disclosure, when harm can still be prevented.

Pricing should reflect the fixed and sunk investments that created the dataset, the value of the access granted, and the effect on future investment. A rule confined to incremental delivery costs tells prospective innovators that regulators may convert successful investments into inputs for rivals without paying for their creation.

Android interoperability should rely on risk tiers. Ordinary invocation presents different concerns than persistent microphone access, screen automation, system controls, aggregation of app data, or background execution. Regulators should permit different treatment of first-party and third-party services when their governance, auditability, revocation procedures, and security responsibilities differ. New capabilities should undergo controlled testing before general release.

Both decisions also need meaningful sunset provisions and empirical review. The test should focus on consumer outcomes. Did users receive better products, lower quality-adjusted prices, stronger security, more innovation, and greater competition among business models? Giving rivals better inputs answers a different question.

When Success Stops Paying

The Commission sees Google’s search data and Android integration as barriers that keep rivals from competing. A dynamic law & economics analysis sees something else as well—the accumulated returns to investment, experimentation, coordination, and risk-taking.

Both views contain part of the story. Data and operating-system access can ease entry, and dominant platforms can use control over important assets to obstruct competition. Yet the mere existence of a valuable advantage does not show that sharing it will improve consumer welfare. It certainly does not show that rivals should receive it at a price largely detached from the cost and risk of creating it.

The July 16 decisions define competition too narrowly, as rival access to the fruits of an incumbent’s investment. They also define innovation too statically, as the number of firms able to use assets that already exist.

Sound innovation policy must ask who will build the next set of assets. A legal regime that promises successful platforms years of administrative redesign, compelled technical support, regulated access prices, and recurring disclosure of sensitive data will not end innovation. It will push firms to design around Europe, delay launches, collect less useful data, integrate less ambitiously, and invest where returns are more secure.

Europe does not lack rules requiring successful platforms to help their competitors. It lacks confidence that success will remain worth the trouble.

The post Europe Wants Tech Champions, Then Makes Them Share the Trophy appeared first on Truth on the Market.

That is a substantial change in how competition law works. The bill would allow CADE to regulate self-preferencing, default settings, interoperability, and data use in advance rather than through ordinary case-by-case enforcement. 

By the government’s estimate, five to 10 companies would be designated, most of them likely American. The stakes therefore extend beyond competition policy. Foreign regulation of U.S. technology firms has become a trade and foreign-policy flashpoint, and Brazil’s proposal could add another point of friction. 

The bill also tests a broader claim now shaping digital regulation around the world. Can traditional antitrust enforcement police dominant platforms, or should governments supplement it with pre-emptive rules modeled on the European Union’s Digital Markets Act (DMA)? 

Brazil offers a revealing test. Its competition authority has recently shown what existing law can accomplish. Settlements with Apple and Google produced DMA-style outcomes, including greater openness in Apple’s iOS operating system, through ordinary case-by-case enforcement. Early evidence from Europe, meanwhile, increasingly points to consumer harm and weaker incentives to innovate rather than the promised surge in competition. 

Brazil’s choice could influence other emerging economies weighing similar regimes. That gives the debate in Brasília significance well beyond Brazil. 

In “Digital Overreach: A Premature Turn to Ex Ante Regulation in Brazil,” Geoffrey Manne, Dirk Auer, and I argued that Brazil neither needs nor would benefit from a new ex ante regime for digital markets. The narrower question here is whether the substitute text now before the Chamber of Deputies improves on the government’s original proposal and how seriously it addresses the criticisms advanced by the International Center for Law & Economics (ICLE) and others. 

The substitute adopts many of the procedural and institutional safeguards we recommended. It leaves the proposal’s doctrinal core intact. That core remains the strongest argument against enacting the bill. 

Fast-Tracked, Then Stalled

The executive branch submitted Bill 4,675/2025 in September 2025. The measure would amend Brazil’s Competition Law, Law 12,529/2011, to let CADE designate large technology platforms as having “systemic relevance” in digital markets and impose “special obligations” on them. 

Two recent procedural developments pushed the proposal closer to a vote. 

The first came March 18, when the Chamber approved an urgency motion, REQ 4612/2025. Under the Chamber’s rules, urgency allows a bill to bypass the standing committees and proceed directly to the plenary floor. That compresses the process into negotiations among party leaders and the bill’s rapporteur, stripping away the procedural checkpoints where legislation is usually tested, amended, delayed, or quietly buried. 

The second came July 8, when the rapporteur, Congressman Aliel Machado of Paraná’s Green Party, filed his report and a substitute text after roughly three months of review. Party leaders had agreed that the report would clear the way for a floor vote, and the Chamber’s website now lists the bill as ready for the agenda. 

The substitute matters because it is, for practical purposes, the text the Chamber will consider. Members may still offer amendments on the floor, but under urgency, the rapporteur’s draft becomes the baseline for every negotiation that follows. 

The timing remains uncertain. An attempt to bring the bill to a vote during the week Machado filed his report failed to win consensus among party leaders. On July 15, the Chamber held its final deliberative session before the winter recess and again left Bill 4,675 off the agenda. 

That leaves a narrow window for action this year. Once lawmakers return to their states, the Chamber is expected to concentrate its legislative work before Brazil’s general elections in October and November 2026. Some expect a vote during the first week of August, after the recess. Brazilian politics advises against putting that date in ink. 

Even Chamber approval would settle only half the matter. Brazil has a bicameral Congress, so the bill would then move to the Senate, where the debate would begin again. 

New Guardrails, Same Bureaucracy

The substitute gives the bill’s new enforcement unit a longer name. It would now become the “Special Superintendence for Systemic Relevance, Free Competition and Consumer Protection in Digital Markets.” The change has no direct legal effect, but it fits Machado’s claim that the bill “regulates markets, not companies.” 

The more consequential revisions limit administrative rulemaking. Article 18-A bars the resolution defining the superintendent’s duties from creating new sanctioning powers, expanding the grounds for designation, or imposing obligations without a statutory basis. CADE’s complementary regulations would be limited to implementing the deadlines and procedures established by law. They could not create additional obligations, sanctions, or grounds for designation. 

That change responds directly to a concern running through ICLE’s analysis. The original bill paired open-ended designation criteria with broad administrative discretion. Brazil’s institutional constraints make that combination especially risky, even if similar powers appear in some of the foreign regimes the bill seeks to emulate. 

The substitute leaves the institutional structure intact. ICLE recommended placing the digital-markets function within CADE’s existing General Superintendence, following the approach used by the European Commission, the United Kingdom’s Competition and Markets Authority (CMA), and Germany’s Bundeskartellamt. Instead, the bill retains a parallel superintendence that duplicates CADE’s architecture and creates room for jurisdictional conflict. 

Worse, the substitute provides no new appropriations to fund it. The bill therefore preserves both the institutional fragmentation and the resource constraints that ICLE identified in the original proposal. 

More Choice, Less Discipline

Article 47-B sets out the new regime’s objectives. The original bill listed three goals: reducing barriers to entry, protecting the competitive process, and promoting freedom of choice. The substitute keeps all three and simply moves “freedom of choice” to the front of the line. 

That reshuffling does not cure the underlying problem. In the “Digital Overreach” white paper and an earlier Truth on the Market post, we argued that Article 47-B departs from the consumer-welfare framework embedded in Brazil’s Competition Law. None of its three objectives contains a limiting principle tied to measurable effects on consumers. 

Barriers to entry may reflect scale economies, quality investments, or network effects that make products better or cheaper. “Protecting the competitive process,” as Herbert Hovenkamp has put it, is little more than a slogan. It is circular, vague, and difficult to measure. 

“Freedom of choice” may be the weakest standard of all. Joshua Wright and Douglas Ginsburg have shown that a choice-based test detaches antitrust from consumers’ actual preferences and the tradeoffs among price, quality, innovation, and variety. It treats more options as inherently better, even when fewer options produce lower prices or better products. 

CADE’s own decisional record points in the same direction. “Freedom of choice” appears only once as a proxy for welfare. Price appears 94 times. 

We recommended making consumer welfare the regime’s explicit objective and limiting principle. The bill’s three stated goals could then serve, at most, as indicators of possible harm. The substitute declines that recommendation. 

That omission creates a deeper statutory tension. The bill would amend Brazil’s Competition Law rather than establish a separate regime. Article 36, Section 1, of that law recognizes that dominance achieved through efficiency and the “natural process” of competition is lawful. Article 88, Section 6, conditions merger approval on benefits reaching consumers. 

An ex ante regime centered on “freedom of choice” fits awkwardly within that framework. Sooner or later, CADE will have to decide which principle governs. 

A Better Gate, the Wrong Key

The designation provisions receive the substitute’s most consequential revisions. These rules determine which companies CADE may classify as having “systemic relevance in digital markets,” Brazil’s term for digital gatekeepers. In an interview, the rapporteur acknowledged that the original test was, to some degree, open-ended. 

The original bill allowed CADE to designate a company based on seven listed characteristics considered “non-cumulatively” and “among others.” The list was therefore nonexhaustive, and any one factor could, in theory, justify designation. The company also had to exceed either R$50 billion in annual global revenue, roughly US$10 billion, or R$5 billion in Brazilian revenue, roughly US$1 billion. 

The substitute replaces that framework with a two-step, closed test. The revenue thresholds become a threshold condition. CADE must then conduct a “joint and reasoned analysis” (análise conjunta e fundamentada) of six listed characteristics. The phrase “among others” is gone.

The substitute also changes how the thresholds may be adjusted. The original allowed ministers to revise them by joint act. The new text indexes them annually to Brazil’s IPCA inflation measure. That removes a political dial and prevents inflation from quietly pulling more firms into the regime, as has happened with Brazil’s merger-review thresholds. 

A new provision also makes explicit what the original bill left unclear. Designation alone would not impose any special obligation. Each obligation would require a separate process and justification. 

Most of these revisions move in the direction ICLE recommended. The “Digital Overreach” white paper criticized the original framework for combining open-ended criteria, politically adjustable thresholds, and a lengthy designation period. The substitute closes the list, fixes the threshold adjustment, requires a reasoned assessment of multiple factors, and, as discussed below, shortens the designation period. 

Two problems remain. First, the six characteristics are still linked by “or,” even though the statute requires a “joint and reasoned” analysis. The drafting points in two directions at once.

Second, some supporters argue that a higher bar would weaken the regime. That objection ignores the cost of designation. A designated firm, along with its entire corporate group, could face years of regulatory obligations, compliance reports, and heightened sanctions risk. 

An error-cost approach weighs the consequences of mistaken intervention against mistaken restraint. In fast-moving digital markets, false positives can be especially costly and hard to reverse. A designation carrying such consequences should therefore require substantial evidence.

The deeper problem remains unchanged. The test still does not require proof of market power. 

The rapporteur’s official release says the bill “does not start from the premise that large companies are a problem.” The statutory test suggests otherwise. ICLE recommended requiring proof of durable market power, consistent with the United Kingdom’s Digital Markets, Competition and Consumers Act (DMCC), which requires “substantial and entrenched market power,” and Germany’s Section 19a, which requires a finding of dominance. 

The substitute instead retains the “systemic relevance” framework. Its six factors remain structural proxies, including multisided business models, network effects, vertical integration, and access to data. Those features may reflect competitive success rather than durable dominance. 

The test also omits the questions a competition economist would ask first. Can new firms enter? Are new technologies disrupting the market, including artificial intelligence? Can users switch to substitutes? 

Designation still applies to the entire corporate group, although the substitute’s service-specific obligations reduce some of the practical breadth. The government also continues to estimate that five to 10 firms would qualify, most of them U.S.-based. That prospect carries obvious geopolitical costs and suggests that company size, rather than demonstrated consumer harm, remains the test’s center of gravity.

On its own terms, the substitute improves the designation test. It is narrower, more predictable, and more demanding. Its central conceptual flaw nonetheless survives.

Fewer Blunt Instruments, More Fine Print

Article 47-E lists the special obligations CADE may impose on designated firms. The substitute reorganizes them into three categories: transparency and reporting duties, positive obligations, and abstention obligations. The changes are substantive, and they cut both ways. 

The clearest improvement concerns merger review. The original bill required designated firms to submit every transaction to CADE, regardless of whether it met the turnover thresholds in Article 88 of Brazil’s Competition Law. That would have created a parallel merger-control regime. 

We argued that the rule was disproportionate. It would capture harmless deals, threaten the startup exit path on which venture investment often depends, and duplicate authority CADE already has under Article 88, Section 7, to call in below-threshold transactions. CADE has recently shown that it will use that power in AI acqui-hire investigations. 

The substitute scales the rule back. Designated firms must now notify the superintendence of below-threshold transactions, following the approach in Article 14 of the DMA. Any review would proceed through CADE’s existing call-in authority. The notice requirement remains blanket rather than risk-based, but the bill drops its most burdensome feature. 

The second improvement is mandatory tailoring. Each special obligation must apply only to specified services, products, or business practices identified in the proceeding. CADE must calibrate the obligation to “the competitive risks associated with each product or service,” and the Tribunal must identify the covered services and set an implementation timetable. 

That change partially adopts ICLE’s recommendation that remedies attach to particular services rather than entire corporate groups. It should prevent the kind of companywide obligations that have made the DMA so costly to administer. The corporate group would still receive the designation, but each obligation would target a particular digital product or service. 

The substitute also strengthens the balancing clause. The original text said CADE “may consider” information security, a firm’s other legal duties, and the functioning of the relevant product or service. The substitute says CADE “shall consider” those factors, as ICLE recommended. 

It also adds two new considerations: innovation, especially for small firms, and the “public interest in the competitive development of digital markets.” The latter phrase sheds little light. Still, mandatory consideration gives CADE a stronger duty to explain how it weighed competing concerns. 

The larger omission remains. The bill still provides no efficiencies defense, a problem discussed below. 

Against those gains, the substitute adds several new mandates. One provision would allow CADE to require “neutral and non-discriminatory” choice architecture, including choice screens, for selecting, installing, or setting third-party products as defaults. Dirk Auer and I have explained to the UK CMA why such remedies can raise consumers’ search costs without producing measurable gains in competition. 

A new abstention rule would also bar firms from worsening conditions for users who exercise rights created by the special obligations. That includes “subversion of user autonomy” through interface design, an anti-circumvention and anti-dark-patterns provision. 

The remaining changes are mixed. The vague ban on “predatory or abusive strategies” survives, though it now applies only where a firm “exploits the situation of dependence” of users. The anti-steering rule grows broader, extending to restrictions “of any other nature.” 

That expansion ignores the role some anti-steering rules play in protecting platform investments in product discovery, trust, and security. CADE’s settlement with Apple showed that those tradeoffs can be assessed through case-by-case enforcement. 

The self-preferencing ban also remains unchanged. Economic research does not support treating self-preferencing as presumptively harmful, and CADE has never based a digital-market conviction on it. 

More Process, Less Ambush

The procedural revisions work best as a package. Together, they make the regime slower, more participatory, and harder to rush. They also respond to ICLE’s concerns about legal certainty, institutional capacity, and rent-seeking. 

The designation period drops from as long as 10 years to six. A firm may seek review after two years if market conditions change significantly. Special obligations may also be revised after two years, and they would take effect within 90 days rather than 60. 

ICLE recommended designation terms of three to five years with mandatory periodic review. Six years, with review triggered only by changed conditions, still exceeds the five-year limits under the UK’s DMCC and Germany’s Section 19a. It is nonetheless a marked improvement over 10 years. 

Participation also expands. CADE must open a 30-day public-comment period when a designation proceeding begins. The consultation period doubles from 15 to 30 days, and the final opinion must expressly address the submissions received. 

A new voluntary-proposal process under Article 87-C allows a designated firm to submit implementation plans, technical parameters, monitoring mechanisms, and timetables. The proposals would not bind CADE, but the agency would have to consider them and explain its response. 

That is a modest step toward the negotiated, case-specific approach ICLE has argued better suits digital markets. CADE has limited experience prescribing product-design remedies. A rule developed with technical input from the firm will often work better than a rigid order that engineers must retrofit after the fact. 

The changes are not uniformly positive. “Any interested party” may still participate, and several public bodies can force CADE to open a proceeding. Referrals from CADE’s Tribunal, its General Superintendence, the Finance Ministry’s Secretariat for Economic Reforms, or any federal body responsible for digital markets or “diffuse and collective rights” trigger immediate proceedings and automatic intervenor status. 

The superintendence may now dismiss unsupported private complaints, which helps. But the bill still needs stronger safeguards against firms and advocacy groups seeking regulatory favors at a rival’s expense. 

The Alternative Becomes an Add-On

The substitute also creates two institutions absent from the original bill. Article 87-K establishes a nonpunitive market-study process. CADE could publish reports to inform designation decisions and evaluate, ex post, whether existing special obligations worked as intended. 

The bill also creates a nonbinding Advisory Council. At least half its seats would go to academics and nonprofit civil-society representatives, with appointment rules meant to reduce the risk of capture. 

ICLE recommended market studies as an alternative to an ex ante regime. The substitute instead adds them to one. Both institutions are sensible on their own. Neither answers the objections to the regime they would support. 

Better Guardrails, Same Destination

The substitute adopts many of the recommendations in the “Digital Overreach” white paper. It closes the designation list, requires reasoned analysis, shortens the designation period to six years, makes clear that designation does not automatically trigger obligations, and requires CADE to tailor each obligation to specific services. It also strengthens the balancing clause, reduces the proposed parallel merger regime to a notice requirement, limits rulemaking, and expands public participation. 

The rapporteur plainly read the criticism and responded to much of it. Four of our central recommendations nonetheless remain unresolved. 

First, the bill still lacks a consumer-welfare anchor. Article 47-B rearranges its objectives but does not reform them. Nothing requires CADE to justify a designation or obligation by showing expected benefits to consumers. 

Second, the bill still lacks a market-power standard. “Systemic relevance” continues to turn largely on size and structural characteristics rather than proven, durable power in a defined market. 

Third, the bill still offers no efficiencies defense. Requiring CADE to consider certain factors improves the procedure, but it does not give a designated firm the right to defeat or narrow an obligation by showing that the challenged conduct produces benefits that outweigh any harm. 

That defense appears elsewhere in Brazilian competition law. Articles 36, Section 1, and 88, Section 6, of Brazil’s Competition Law recognize efficiencies as legally relevant. The new regime does not. As a result, the same conduct could be lawful under the statute’s general provisions because it benefits consumers, yet prohibited under the digital-markets regime because efficiency provides no defense. The substitute leaves “economic justification” undefined, so the inconsistency remains. 

Fourth, the bill still requires no regulatory impact assessment, even though Brazil’s Economic Freedom Act, Law 13,874/2019, requires such analysis for regulatory measures of far less consequence. 

The substitute therefore produces two conclusions. It is a serious effort to add guardrails, and on several procedural dimensions, Brazil’s proposal is now more constrained than the DMA. 

Those guardrails also expose the central defects more clearly. CADE could still designate a firm without proving durable market power. It could still impose a tailored obligation that the firm cannot challenge by showing that the conduct benefits consumers. 

A well-run process is no substitute for a sound legal standard. Better machinery only makes the missing metric harder to ignore. 

A Global Trend, or Just a Rerun?

The rapporteur has also pointed to developments abroad as evidence that the international trend still favors ex ante regulation. Beyond the European Union, United Kingdom, and German regimes already in force, he cites the June 10 reintroduction of the American Innovation and Choice Online Act (AICOA) by Sens. Amy Klobuchar (D-Minn.) and Chuck Grassley (R-Iowa). 

Context matters. This is AICOA’s third appearance after failing in the 117th and 118th Congresses. As my colleague Daniel Gilman put it, the latest version “is not so much a fresh start as a sequel nobody ordered.” It arrived with four original co-sponsors, in an election year, and with no clear path that its predecessors lacked. 

Reintroduction alone does not amount to regulatory momentum. It is a thin basis for claiming that Brazil would be joining a global movement rather than racing ahead of one. 

Nor does AICOA offer a model worth copying. Geoffrey Manne observed that the new version “fails to fix the bill’s central legal and economic flaws—and in several ways makes them worse.” It still identifies covered firms, now called “systemically important platforms,” through formulas based on revenue and user reach, without requiring proof of market power. 

Herbert Hovenkamp’s critique tracks the central problem with Article 47-C. AICOA selects firms “on the basis of raw size rather than market power,” moving competition law away from anticompetitive conduct and its effects and toward structural traits that may have no connection to consumer harm. 

The same criticism applies to Brazil’s bill. Brazilian law already recognizes the contrary principle. Under Article 36, Section 1, of Brazil’s Competition Law, a company does not violate antitrust law merely by becoming large or dominant through efficiency. 

The rapporteur has shown that the bill can be drafted more carefully. The substitute adds real procedural safeguards, and further revisions could improve it again. 

But better drafting cannot answer the threshold question. Brazil still has not shown why it needs this regime in the first place. 

A sharper scalpel does not justify an unnecessary surgery.

The post Brazil’s Gatekeeper Bill Gets a Tuneup, Not a Rethink appeared first on Truth on the Market.

Regulators and academics warn that a small group of firms—including Amazon, Anthropic, Google, and OpenAI, with Microsoft and Meta sometimes added—will dominate generative artificial intelligence. Their advantages in computing power, capital, data, and distribution will harden into market power, shut out challengers, and concentrate control over a technology reshaping the economy. 

Public agencies have embraced this account. The U.S. Federal Trade Commission (FTC), U.S. Department of Justice (DOJ), U.K. Competition and Markets Authority (DMA), and European Commission advanced it in their Joint Statement on Competition in Generative AI, which I have discussed before.

Academics have raised similar concerns, even as the sector continues to grow quickly. Much of their attention centers on “GAMMA”—Google, Amazon, Microsoft, Meta, and Apple—and those firms’ control over critical inputs such as data and computing capacity. The fear is that these advantages could produce an “AI oligopoly.” 

That theory has given regulators a reason to act early. In the European Union, it has shaped Digital Markets Act (DMA) specification proceedings, Article 102 investigations, and emergency interim measures. In Brazil, it has pushed merger review beyond mandatory notification thresholds to reach AI partnerships. In Italy, it has prompted proceedings over Meta’s integration of AI into WhatsApp before regulators established any consumer harm. 

As I argued in an earlier post, these interventions follow the same logic. Regulators treat a plausible theory of harm as enough to justify immediate action, while giving limited weight to the safeguards that ordinarily discipline competition enforcement. That approach risks suppressing the very practices through which AI firms compete, including integration, partnerships, and the use of existing distribution networks. 

Yet the larger flaw appears even earlier in the analysis. The story remains almost entirely Western. Regulators cast GAMMA, OpenAI, and Anthropic as the firms to contain, then focus on their relationships with one another and with U.S. and European complementors, meaning companies whose products or services increase the value of another firm’s offering. 

Meanwhile, one of the fastest-growing sources of competitive pressure in global AI barely appears in market-definition exercises, foreclosure theories, or claims that power in older markets will carry over into AI. 

China is missing from the case file. 

The Competitor Regulators Forgot

Chinese models accounted for roughly 1% of the global generative AI market in late 2024. By the end of 2025, their share had climbed to about 15%, according to data reported by Nikkei. 

The Chinese market has coalesced around about 10 serious providers, each offering frontier-class models. DeepSeek and Alibaba’s Qwen are the best known abroad, but the field also includes Zhipu’s GLM-5, Moonshot AI’s Kimi, MiniMax’s M-series, ByteDance’s Doubao and Seed models, Baidu’s ERNIE, Tencent’s Hunyuan, StepFun’s Step series, and Xiaomi’s MiMo. Kimi’s latest version reportedly beats Fable 5 on some benchmarks, while MiMo has emerged as a surprise volume leader on global developer platforms. 

Several of these models remain primarily domestic. Baidu’s ERNIE, Tencent’s Hunyuan, and ByteDance’s Doubao are closely integrated into Chinese search, messaging, and device products, and remain difficult to access abroad. 

Others follow a very different model. DeepSeek, Qwen, GLM-5, and Kimi K2.5 distribute their model weights under permissive licenses, including MIT- and Apache-style terms. “Open weights” means developers can download the underlying model parameters, adapt them, and deploy the resulting systems on their own infrastructure. A developer in Europe, the United States, or Latin America can use these models without opening an account, obtaining a Chinese phone number, or relying on a Chinese server. 

Qwen alone has surpassed 700 million downloads on Hugging Face, overtaking Meta’s Llama as the world’s most downloaded AI-model family. Developers have built more than 113,000 derivative models from its checkpoints, or saved versions of the model used as starting points for further training. These products do not need to find Western distribution. They already have it. 

The performance gap has narrowed just as quickly. Stanford University’s 2026 AI Index Report estimates that the gap between the best U.S. and Chinese models has fallen to 2.7 percentage points, compared with 17.5 to 31.6 points in May 2023. U.S. firms attract 23 times as much private AI investment as Chinese firms, yet lead in model performance by less than 3 percentage points. Among leading open-source models, U.S. and Chinese systems have repeatedly traded the top spot on major benchmarks since early 2025. 

None of this appears in the public record of the European Commission’s proceedings against Meta; the Brazilian Administrative Council for Economic Defense’s (CADE) referrals involving Amazon, Microsoft, and Google AI partnerships; or the DMA specification proceedings concerning Alphabet. 

The usual caveat applies. Market definitions depend on the facts of each case, and regulators tailor theories of harm to particular conduct. Some proceedings may have sound reasons to exclude Chinese models. 

The broader pattern is harder to defend. No major AI enforcement action in Europe, the United States, or Latin America has publicly confronted the possibility that some of the strongest competitive pressure in global AI comes from firms beyond those regulators’ reach. Academic debate often makes the same omission. Chinese AI scarcely exists in either account. 

That omission looks stranger against the growing concern in trade policy. Chinese firms aim to export lower-cost alternatives to Western graphics-processing units (GPUs), the chips used to train and run AI models, while spreading open-source models abroad to build long-term dependence on Chinese technology. Chinese models have already gained users among some of the largest U.S. companies. 

Competing Outside the Antitrust Playbook

Chinese AI firms are pursuing a competitive strategy that differs sharply from the proprietary, vertically integrated model favored by leading U.S. companies—and from the market structure most antitrust tools assume. 

OpenAI, Anthropic, and Google DeepMind develop frontier models, control access through application programming interfaces (APIs), and build applications, distribution deals, and enterprise services around them. They generally treat scale, computing power, and proprietary training data as core competitive advantages. Much Western antitrust enforcement in AI rests on the premise that regulators must police those advantages, especially when they combine with the distribution networks of established platforms. 

Many Chinese firms compete differently. They rely more heavily on open weights, inexpensive fine-tuning, and state-supported distribution. DeepSeek releases V3.2 and V4 under the MIT license, with model weights available for commercial use. Alibaba’s Qwen family uses a hybrid approach. Its midrange models, up to 35 billion parameters, remain available under the Apache 2.0 license, while its most capable models have moved toward proprietary access. That shift resembles, with some delay, the drift toward closed systems among U.S. frontier labs. 

For the open tier, ubiquity is the competitive weapon. A model that is free to download and cheap to run can spread through adoption rather than through controlled access. 

That difference creates two problems for current enforcement. 

First, diffusion through open-source models does not fit neatly within the remedies regulators now favor. Behavioral restrictions, data-sharing mandates, and interoperability rules assume identifiable firms, proprietary products, and gatekeeping intermediaries. Those tools have little purchase on a model family downloaded 700 million times and embedded in derivative applications across multiple jurisdictions. 

Second, open models weaken some of the market-power theories behind current cases. Consider the European Commission’s DMA proceeding extending search-data-sharing obligations to AI chatbots. The concern is that Google could use its dominance in search to secure dominance in AI assistants. 

That theory depends on Google controlling an input that rivals cannot obtain elsewhere. Yet Qwen, DeepSeek, and, more recently, Kimi have gained users worldwide without access to Google’s search data. They did so by offering capable models at low cost. Their growth complicates the Commission’s foreclosure theory. 

Different business models do not place Chinese firms in a separate market or make them relevant to every antitrust dispute. Open-source models can still discipline leading providers, even when their licensing, distribution, and revenue models differ. 

The competitive pressure also extends beyond open source. Chinese firms offer leading models through inference APIs, which allow developers to send requests to remote models without operating the underlying infrastructure. Those services remain accessible worldwide, though they carry the security and privacy risks associated with Chinese-operated systems. 

Some technically open models still require enormous computing resources. Zhipu’s GLM-5, for example, uses a 744-billion-parameter mixture-of-experts architecture and is available under the MIT license. Few developers can run it themselves. Most reach it through third-party inference services such as OpenRouter. 

That still counts as competition. A model need not run on a laptop to pressure incumbent providers. If developers can reach it through the same services they already use, it competes on price, capability, and availability with leading U.S. models. 

Competition also will not always pit one frontier model against another. Many production systems use several models for different tasks. A larger controller model may handle open-ended reasoning, while smaller, specialized models parse inputs, format outputs, and route tool calls. 

Those routine tasks rarely require frontier-level performance. Firms can assign them to fine-tuned small language models at far lower cost. Recent analysis suggests that using a frontier model for the roughly 30% of tasks requiring advanced reasoning and a smaller model for the remaining 70% can cost about one-tenth as much as sending every task to a large model. 

Smaller models are not perfect substitutes for frontier systems. They still create meaningful competitive pressure. Developers and enterprises seeking to control costs can reduce their dependence on any single frontier-model provider by combining larger systems with smaller Chinese or non-Chinese alternatives. 

The Market Has Noticed, Even If Regulators Haven’t

At least one major generative AI company has noticed the competition from China. In a May 14 blog post, Anthropic described two possible paths through 2028. Under the first, the United States and allied democracies retain a 12- to 24-month lead in frontier AI. Under the second, China closes the gap and reaches parity. 

Anthropic’s analysis reflects its assessment of current competitive trends. Without changes to export controls, computing policy, and investment priorities, the company believes China could erase the remaining gap within two years. Dario Amodei made a similar argument in his earlier essay, “On DeepSeek and Export Controls.” He described DeepSeek as a genuine competitive challenge, though one that some observers had overstated, and argued that stricter export-control enforcement was necessary to preserve the U.S. lead. 

Microsoft’s conduct offers an even clearer test. According to Axios, the company is considering a fine-tuned version of DeepSeek V4, hosted on Azure, as a cheaper alternative to the OpenAI and Anthropic models that power Copilot Cowork. Charles Lamanna, Microsoft’s executive vice president for Copilot, gave a simple reason. Some enterprise users perform hundreds of tasks each week, and “the costs can go very high.” 

The price advantage of Chinese open-source models is large enough that Microsoft, an OpenAI investor and close commercial partner, may route enterprise workloads through a Chinese-origin model. Other companies have already made that choice: 

Lindy, a San Francisco-based company that builds AI work assistants, recently made a switch from Anthropic models to DeepSeek, according to its founder Flo Crivello, who announced the move on X in June. Crivello said the switch saved the firm millions of dollars. “You don’t need God to write your email,” he said on tech news show MTS. “If you can get those lower tiers of intelligence for a tenth of the price, it would be foolish not to do it.”

All three major U.S. cloud providers now offer Chinese open-source models through managed application programming interfaces. Amazon Bedrock hosts DeepSeek, Qwen, Kimi, MiniMax, and GLM. Google Cloud Vertex AI offers DeepSeek and Kimi through fully managed serverless interfaces. Azure AI Foundry includes DeepSeek and Kimi in its model catalog. 

The competitive pressure is already reaching the enterprise market. Proprietary U.S. companies must now compete with cheaper Chinese models distributed through their own cloud services. Antitrust analysis that ignores those models is describing a market that its largest participants no longer recognize. 

Investors appear to recognize the threat as well. DeepSeek raised more than $7 billion last month, the largest financing round in Chinese AI history, at a valuation above $50 billion. The terms were unusual. Investors accepted a five-year lockup and no voting rights, while only China’s National Artificial Intelligence Industry Investment Fund invested directly. Even on those terms, investors committed capital on a scale that suggests they expect DeepSeek to remain a serious global competitor. 

Regulating One Side of the Race

Western antitrust enforcement imposes compliance costs, procedural burdens, product-design limits, and uncertainty on AI companies operating in the European Union and the United States. Chinese AI firms face few comparable constraints when competing in those same markets. 

That asymmetry does more than burden a particular group of companies. It can distort competition and leave consumers worse off. 

Judge Frank Easterbrook’s error-cost framework warns that false positives—mistakenly condemning conduct that helps competition—can be especially costly in young markets. Competitive conditions remain unsettled, and markets may correct themselves faster than regulators can. Premature intervention becomes even riskier when foreign rivals are ready to capture the business lost by firms constrained through antitrust remedies. 

The short-term effects may look attractive. Smaller firms may enter and gain market share. The harder-to-see costs may include weaker integration, less innovation, and lower investment in computing capacity, data, and other critical inputs. 

Consider the DMA requirement that Google share search data on fair, reasonable, and nondiscriminatory (FRAND) terms. The European Commission expects that access to help rival search engines and AI services enter the market. 

That theory assumes a contest between a dominant U.S. incumbent and smaller American or European challengers. The market now looks different. Chinese open-source models have accumulated hundreds of millions of downloads, and Microsoft is considering them for enterprise use. Mandated access to Google’s search data may benefit firms far beyond the European rivals the Commission had in mind. 

More broadly, Western incumbents bear growing regulatory costs while Chinese competitors operate beyond the reach of those rules. Most competition theory pays little attention to that imbalance. 

Restrictions that weaken the product quality or integration advantages of Google, Meta, or Anthropic may do little for consumers. They may instead shift market share toward firms outside the regulatory perimeter while leaving users with worse products. 

The Market Is Bigger Than the Case File

None of this means Chinese AI competition is uniformly benign, that Western regulators should abandon enforcement, or that current antitrust concerns are imaginary. Competition policy still has a role in policing genuine foreclosure, preventing exclusionary access terms, and blocking mergers that harm competition. 

Enforcement should rest on evidence, sound economic theory, and a complete account of the market. If Chinese open-source models are growing quickly enough that leading U.S. firms treat them as a serious strategic threat, regulators cannot analyze AI competition as a contest among Western incumbents alone. 

Theories that assume dominance in search, cloud computing, or social media will automatically translate into AI dominance must account for adoption data showing Chinese models gaining global users at extraordinary speed. Claims that AI chatbots risk being shut out of distribution must confront the fact that some of the world’s most downloaded models are available free to developers almost anywhere. 

Two caveats temper the argument. First, Chinese adoption data are often difficult to verify independently. Qwen’s reported 700 million Hugging Face downloads, for example, comes largely from company announcements and cannot be easily audited. Regulators should consider such figures without treating them as gospel. 

Second, geopolitics may divide the global AI market. U.S. semiconductor export controls, data-localization rules, and enterprise security policies already discourage Chinese software in some regulated industries. Chinese and Western systems may increasingly serve separate markets rather than compete directly. If that division deepens, some of today’s open-weight competition may become geographically limited. 

Neither caveat resolves the central problem. Even heavily discounted, the available figures point to Chinese competition growing at a pace and scale that current enforcement has barely addressed. Market separation would also reflect choices made by governments, firms, and regulators. It does not justify defining today’s market as though Chinese providers were absent. 

Antitrust authorities are right to scrutinize AI markets. They should start by looking at the whole market.

The post The Missing Rival: China and the Limits of AI Antitrust appeared first on Truth on the Market.

That gap between scientific promise and clinical use tells us more about American innovation policy than most congressional hearings ever will.

At one such hearing earlier this week, members of the Senate Judiciary Committee returned to one of the most consequential and least understood questions in innovation policy. Which inventions qualify for patent protection? The hearing focused on S. 1546, the Patent Eligibility Restoration Act (PERA), Congress’ latest attempt to resolve more than a decade of confusion over Section 101 of the Patent Act.

The debate turned on statutory text, Supreme Court precedent, and competing theories of patent policy. Senators heard about abstract ideas, laws of nature, diagnostic methods, and judicial exceptions to patent eligibility. Those questions matter, but they can quickly drift away from the inventions and patients affected by them.

To see the stakes, the committee need look no further than a trauma bay at UPMC Presbyterian Hospital in Pittsburgh.

Physicians there can now use a blood test to help determine whether an adult with a suspected traumatic brain injury needs a CT scan. Instead of relying only on visible symptoms or sending every patient for costly imaging, doctors can measure two proteins released into the bloodstream after brain injury. Those proteins are glial fibrillary acidic protein (GFAP) and ubiquitin carboxyl-terminal hydrolase L1 (UCH-L1).

The test gives physicians objective biological evidence in one of emergency medicine’s most difficult diagnostic settings. Until recently, it remained a scientific aspiration.

Its 20-year path to patients captures both the strengths of the American life-sciences system and the institutional burdens that can slow it down. Abbott’s i-STAT TBI test emerged through university research, federal funding, patents, startups, venture capital, technology transfer, clinical trials, strategic acquisitions, Food and Drug Administration (FDA) review, and entrepreneurial judgment. It also encountered the uncertainty created by judicial decisions that narrowed patent eligibility for diagnostic inventions.

The history of the test reaches well beyond diagnostics. It shows how people and institutions make costly decisions under profound uncertainty to create knowledge that did not exist before. It also shows why innovation policy should be judged by whether it helps today’s discoveries become tomorrow’s treatments.

The Long Road Out of the Lab

The story begins in a university laboratory.

More than two decades ago, neuroscientist Kevin Wang and his colleagues at the University of Florida began studying whether proteins released after a traumatic brain injury could help physicians detect damage that clinical observation often missed. The National Institutes of Health (NIH) and the Department of Defense (DOD) funded much of the work, reflecting the importance of brain injuries to civilian medicine and to military personnel returning from combat.

At that point, there was no commercial product, no FDA-cleared test, and no global distribution plan. There was a scientific hypothesis that Wang and his colleagues believed might someday improve patient care.

Most promising discoveries never get much further. Many disappear into the “valley of death,” the long, expensive, and uncertain period between laboratory research and commercial use. A technology must survive further study, product development, clinical validation, regulatory review, manufacturing, distribution, and repeated rounds of financing before it reaches a patient.

Money is only part of the problem. The process also depends on entrepreneurial judgment under genuine uncertainty. Scientists cannot know whether early findings will survive later experiments. Investors cannot know whether years of research will yield a viable product. Entrepreneurs must commit scarce resources without knowing whether regulators will approve the technology, physicians will adopt it, or competitors will produce something better first.

Ludwig von Mises described this kind of decision-making as purposeful human action directed toward an uncertain future. As I have argued, innovation does not emerge predictably from a machine-like process. It depends on people who interpret incomplete information, imagine possible futures, and act before the outcome is known.

That view differs sharply from policy accounts that treat innovation as the automatic result of greater research spending, industrial policy, or government planning. Those accounts assume that more inputs will produce more outputs. They leave little room for judgment, error, adaptation, or discovery.

The Austrian tradition begins with human choice. People form expectations, weigh risks, and commit resources today in hopes of creating something valuable tomorrow. Entrepreneurship is therefore central to innovation because every meaningful advance requires someone to act before success can be measured.

The traumatic-brain-injury (TBI) blood test illustrates that process at every stage of its development.

Patents Make Strange Bedfellows

Wang and his colleagues knew that academic research alone would not bring the test to patients. In 2002, they helped establish Banyan Biomarkers to turn the university’s discoveries into a commercial diagnostic.

That move from laboratory to startup reflects a familiar division of labor in American innovation. Universities specialize in basic research. Startups test, refine, and develop new technologies. Large firms bring manufacturing capacity, regulatory expertise, and global distribution. Each does something different, and none does everything equally well.

This specialization matters most in knowledge-intensive industries, where scientific discovery and mass production demand very different skills. Jonathan Barnett calls research-focused organizations “idea factories.” They concentrate on developing technology, then use intellectual property rights and contracts to work with firms that possess the capabilities needed for commercialization.

Banyan fits that model closely. It lacked Abbott’s manufacturing and distribution network. Abbott could not recreate two decades of specialized biomarker research on demand. Each possessed knowledge and capabilities the other lacked. Patents allowed them to combine those assets rather than duplicate them.

That role complicates the familiar claim that patents mainly confer monopoly privileges and suppress competition. In biotechnology and the life sciences, patents often make cooperation possible. They give researchers enough legal protection to disclose discoveries, negotiate licenses, attract investment, and work with commercial partners without surrendering the invention the moment they reveal it.

Kenneth Arrow identified the underlying problem decades ago. A prospective buyer often must learn an idea before deciding what it is worth. Once the buyer knows the idea, though, the seller may have little left to sell. Intellectual property partly resolves that paradox by allowing innovators to share information while retaining enforceable rights.

Barnett carries the argument further. Strong intellectual property rights allow researchers and manufacturers to specialize according to comparative advantage. Research organizations need not build factories and distribution networks. Manufacturers need not reproduce every scientific discovery themselves. Licensing and collaboration let each side contribute what it does best, reducing both development costs and time to market.

Abbott’s 2019 acquisition of Banyan therefore marked more than the closing of a corporate deal. It capped a decadeslong process in which universities, startups, investors, clinicians, and established firms supplied knowledge, capital, and organizational capabilities that no single institution possessed.

Friedrich Hayek argued that economic knowledge exists in dispersed fragments rather than in one mind or institution. Markets coordinate those fragments by allowing people with partial knowledge to cooperate through exchange. The TBI test offers a vivid example.

No central planner designed its path. No agency coordinated every stage. Researchers, entrepreneurs, investors, clinicians, patent lawyers, technology-transfer professionals, regulators, manufacturers, physicians, and patients each contributed specialized knowledge. Institutions that support property rights, contracts, investment, and exchange allowed those contributions to become a usable diagnostic.

The result is a scientific achievement made possible by entrepreneurial coordination.

The Government’s Patent Attic

The TBI blood test also depended on one of the most consequential innovation laws Congress has enacted: the Bayh-Dole Act of 1980.

Bayh-Dole allowed universities to retain ownership of inventions developed through federally funded research. That familiar description is accurate, but it understates both the problem Congress faced and the importance of the remedy.

Before Bayh-Dole, ownership of federally funded inventions usually followed the funding rather than the inventor. Universities often agreed that discoveries produced with federal support would belong to the government. The arrangement may have looked tidy on paper. In practice, tens of thousands of government-owned patents accumulated, while relatively few were licensed or commercialized.

Promising discoveries stalled because the institutions best positioned to develop them neither owned them nor had strong incentives to invest in them. The defect lay in the structure itself.

The American patent system traditionally linked invention, ownership, and voluntary exchange. Inventors created new knowledge, property rights allowed them to control and transfer it, and markets helped move it into wider use. The Framers reflected that understanding in the Constitution, which empowers Congress to secure authors and inventors exclusive rights for limited periods to promote science and the useful arts.

Bayh-Dole restored that link for federally funded university research. Universities could retain title to faculty inventions, negotiate licenses, form startups, attract private investment, and partner with firms that possessed the manufacturing, regulatory, and distribution capabilities needed to reach patients.

Its success came largely from removing a government-created barrier to commercialization. Though Bayh-Dole is a federal statute, it worked by giving universities and inventors greater control over how discoveries moved into private development.

That distinction matters. Effective reform often requires clearing away rules that block productive exchange. Contemporary innovation policy too often moves in the opposite direction, layering new interventions atop the unintended consequences of older ones.

The TBI test’s subsequent history shows how costly that habit can become.

The Cost of Taking Our Time

The successful commercialization of the TBI blood test deserves celebration. Its 20-year journey does not.

Some of that delay reflects the difficulty of medical innovation. Clinical testing takes time. Regulators must protect patients. Manufacturing sophisticated diagnostics at commercial scale is demanding. None of that resolves the central question. Should a test that can improve traumatic-brain-injury diagnosis really take two decades to reach routine clinical use?

Innovation policy rarely asks. Regulators and lawmakers focus far more on the dangers of moving too quickly than on the costs of moving too slowly. Agencies are judged for visible failures, such as approving an unsafe product. They receive little credit for benefits that arrive sooner because a needless delay was avoided.

Yet delay carries real costs. Patients continue to receive less effective care. Physicians make decisions with poorer information. Researchers and entrepreneurs spend time and money answering regulatory concerns instead of advancing the technology. Investors keep funding facilities, equipment, and personnel without knowing when, or whether, a return will come.

Many discoveries never survive the process.

Those losses rarely appear in regulatory accounting because they consist mostly of opportunities forgone. A regulator cannot identify the patient who might have benefited five years earlier. An agency cannot count the startup that was never formed because commercialization looked too uncertain. A policymaker cannot see the investor who shifted capital to a safer project after watching another technology stall.

Hayek’s account of dispersed knowledge applies with particular force. The information needed to assess innovation is spread among scientists, entrepreneurs, physicians, investors, manufacturers, and patients. Each knows something the others do not. No regulator can collect all of it, especially when the missing knowledge concerns futures that delay prevents from ever occurring.

That is why the TBI test’s 20-year path should inspire concern alongside admiration. If this is one of the success stories, the more troubling question concerns the inventions that never made it out.

Whose Risk Is It, Anyway?

These concerns lead to one of innovation policy’s hardest questions. Who should decide when a new technology is safe and effective enough for broader use?

The conventional answer is expert regulators. The Austrian answer begins with a caveat. Expertise matters, but it is always incomplete. Even the most capable regulator possesses only part of the knowledge needed to make decisions for millions of patients with different conditions, values, and tolerances for risk.

Medicine makes that limitation plain. Safety and efficacy are not absolute measures. Both require judgments about tradeoffs under uncertainty, and those tradeoffs vary across patients.

A healthy college athlete with a suspected concussion may tolerate less diagnostic uncertainty than a 75-year-old patient with several serious conditions. A patient with terminal pancreatic cancer may accept risks that someone seeking treatment for a minor illness would reject. Those choices reflect life expectancy, quality of life, family circumstances, personal values, and appetite for risk.

No formula can fully capture those differences. Patients possess that knowledge, often with help from physicians who understand both the evidence and the individual case. A centralized regulator, however skilled or well intentioned, cannot know every patient’s preferences.

Former Sen. Ben Sasse’s diagnosis with metastatic pancreatic cancer illustrates the point. After doctors reportedly told him that he might have only months to live, Sasse gained access to daraxonrasib through the FDA’s expanded-access program. The drug had shown promise in clinical trials and had already received Breakthrough Therapy and Orphan Drug designations, though it had not yet won full approval.

Sasse reportedly experienced a sharp reduction in tumor volume. His case does not prove that every experimental therapy should become immediately available, or that regulatory oversight serves no purpose. It shows that patients facing radically different circumstances can rationally value uncertainty in radically different ways.

Someone facing probable death within months may accept risks that a healthy regulator or lawmaker would refuse. That judgment is not irrational. It reflects consumer sovereignty.

Mises argued that market economies ultimately answer to consumers rather than producers or planners. In medicine, those consumers are patients working with physicians who understand the science and the patient’s circumstances.

The current FDA model asks regulators to choose the proper balance among safety, efficacy, speed, uncertainty, and opportunity cost for millions of different patients. The knowledge needed to make that choice does not exist in one place.

Policymakers should therefore consider making accelerated pathways more broadly available to patients facing the gravest risks and fewest alternatives. Earlier access could reduce development costs, shorten the path to market, and allow adoption to proceed gradually among patients with the strongest reasons to accept uncertainty.

That approach could also improve safety monitoring. Rare risks often emerge only after a treatment reaches real-world patients, including those excluded from clinical trials. Gradual uptake among patients with high expected benefits could reveal those risks before widespread use, rather than concentrating adoption during a heavily marketed launch.

How to Patent Around a Court

If Bayh-Dole shows how removing a legal distortion can encourage entrepreneurial discovery, the Supreme Court’s recent patent-eligibility decisions show how creating one can impede it.

The shift began with Mayo Collaborative Services v. Prometheus Laboratories in 2012, followed by Association for Molecular Pathology v. Myriad Genetics and Alice Corp. v. CLS Bank. Together, those decisions transformed the law governing which inventions qualify for patents.

Section 101 of the Patent Act had long served as a “coarse filter,” reflecting Congress’ broad definition of patentable subject matter. The Court instead made eligibility turn increasingly on judge-made exceptions for laws of nature, natural phenomena, and abstract ideas.

The result has been years of uncertainty for inventors, patent examiners, investors, district courts, and even the U.S. Court of Appeals for the Federal Circuit, which hears patent appeals. PERA’s congressional findings acknowledge the confusion and inconsistency that followed.

The TBI blood test shows why that uncertainty carries practical consequences. A patent claiming only the discovery that elevated levels of GFAP and UCH-L1 correlate with traumatic brain injury might face rejection as an attempt to patent a law of nature. Patent prosecutors responded as people usually do when the rules change. They adapted.

The resulting patent portfolio claims far more than the biomarkers themselves. It covers assays, antibodies, cartridges, point-of-care systems, patient-management methods, imaging decisions, and integrated diagnostic processes.

Earlier diagnostic patents often stated the invention directly. Later patents increasingly resemble elaborate commercialization blueprints constructed to survive a doctrine that may not protect the underlying scientific discovery.

Patent lawyers did not suddenly become more inventive, nor did scientists become more acquisitive. Courts changed the rules, and innovators changed their behavior in response. Much of contemporary innovation policy misses that simple point.

The Court Changes the Rules, Everyone Else Adapts

One of Mises’ most durable points is that people do not passively submit to institutional design. They adjust their behavior when incentives change. The point sounds obvious until policymakers forget it.

The Supreme Court likely believed it was addressing a genuine problem in Mayo. Whether one agrees with the decision or not, the justices’ concern was clear. Broad patents on natural laws might impede scientific progress rather than promote it.

The difficulty lay in the consequences the Court could not predict. It could change legal doctrine, but it could not know how inventors, entrepreneurs, universities, venture capitalists, technology-transfer offices, patent lawyers, and established firms would respond once the new incentives spread through the patent system.

That knowledge did not yet exist.

Hayek’s “The Use of Knowledge in Society” is often described as an argument about dispersed information. Its deeper point concerns knowledge that emerges only through action. The information needed for economic coordination exists in fragments, scattered among people whose plans and expectations change as circumstances change.

No central decision-maker can collect all of that knowledge because much of it appears only after people begin adapting.

The Supreme Court therefore took on a task no court could complete. It might predict how judges would apply a new patent-eligibility test. It could not predict how inventors and firms would redesign claims, alter investment decisions, abandon research, or build new legal strategies around “laws of nature,” “natural phenomena,” and “abstract ideas.”

That is no indictment of the Court’s intelligence. It is a limit on what any institution can know.

The Thicket Thickens

Contemporary patent policy contains a sizable irony. Critics now argue that pharmaceutical “patent thickets,” portfolios of overlapping patents covering a single product, impede competition. The recently introduced Eliminating Thickets to Increase Competition (ETHIC) Act rests on that premise and would limit the assertion of patents linked through terminal disclaimers in pharmaceutical litigation.

The immediate question is not whether the bill is sound. A more useful question comes first. Why did these increasingly complex portfolios emerge?

The standard answer blames firms seeking to prolong their “monopolies” by abusing the patent system. That may explain some conduct. The TBI blood test suggests another cause.

Suppose diagnostic claims had remained reliably eligible for patents under Section 101. Innovators might have relied on a smaller number of properly scoped claims rather than assembling portfolios covering diagnostic methods, antibodies, cartridges, devices, software, manufacturing techniques, and clinical workflows.

Patent prosecution is expensive, and maintenance fees accumulate. A limited portfolio, updated through continuation applications as the technology and market develop, is often simpler and cheaper.

Continuation practice serves a legitimate purpose. It allows patent claims to evolve as inventors learn more about commercial uses, competitors, and practical versions of an invention. The harder question is whether uncertainty over diagnostic patents now pushes innovators to file more continuations, divisional applications, and related patents than they otherwise would.

If so, some patent thickets may reflect a rational response to legal uncertainty. When courts make direct protection for a diagnostic discovery unreliable, innovators spread protection across multiple narrower patents.

That possibility changes the policy story. At least some patent thickets may be consequences of an earlier intervention rather than independent defects in the patent system. Innovators adapted to the incentives created by judicial doctrine.

The cycle is hard to miss. One intervention changes incentives. Firms respond. Policymakers then cite those responses as grounds for another intervention.

The Cure for the Cure

Nearly a century ago, Mises described this pattern in his critique of interventionism. Government action often produces results policymakers did not expect. Instead of revisiting the original policy, they treat the resulting distortions as proof that another intervention is needed.

The cycle then repeats. Each new rule tries to repair the damage caused by the last one while leaving the original source of the problem intact.

American innovation policy increasingly follows that script. Federal ownership rules weakened incentives to commercialize publicly funded research. Bayh-Dole removed much of that barrier, and commercialization accelerated.

The Supreme Court’s patent-eligibility decisions then created new uncertainty. Innovators responded by changing how they drafted claims and structured patent portfolios. Lawmakers now condemn those more complicated portfolios as abuse and propose limiting their use.

Enter the ETHIC Act.

Whether the bill becomes law is almost secondary. The larger pattern is what matters. One intervention distorts incentives. Innovators adapt. Policymakers then treat that adaptation as a new defect requiring still more intervention.

The deepest cost is that each new layer makes the original mistake harder to see.

No One Has Neutral Incentives

Public Choice economics adds another question that innovation policy often skips. Why assume judges, regulators, legislators, or agencies have better incentives than entrepreneurs, physicians, investors, or patients?

Expertise alone cannot answer it. Expertise matters, but knowledge and incentives are different. Scientists, physicians, patent examiners, judges, and entrepreneurs each know different things. The harder question is whether any institution possesses enough information, and the right incentives, to direct entrepreneurial discovery better than a decentralized process.

Public Choice scholars have long argued that government officials respond to incentives just as private actors do. Agencies seek larger budgets. Legislators answer to constituencies. Courts favor rules they can administer. Universities compete for grants. Companies seek profits. Venture capitalists seek returns.

No one occupies a neutral perch above the system. Everyone responds to incentives.

Recognizing that does not require cynicism. It requires humility, especially about the ability of any institution—even the Supreme Court—to predict how millions of purposeful actors will respond when the rules change.

Congress, Clean Up the Court’s Mess

That brings us back to the July 14 Senate hearing. Congress should not have to repair patent-eligibility doctrine, but it does.

The current framework is unworkable. Lower courts struggle to apply standards that judges themselves often describe as incoherent. Inventors, entrepreneurs, and investors remain uncertain about what qualifies for protection, especially in diagnostics and computer-implemented inventions. That uncertainty has weakened U.S. innovation in fields where the country can least afford hesitation.

The Patent Eligibility Restoration Act reflects an important institutional judgment. Congress, not the judiciary, bears responsibility for defining patent eligibility. PERA’s greatest contribution may be its recognition that the courts created exceptions and confusion that Congress never enacted.

I am less persuaded that Congress should replace one complicated framework with another. A simpler and more constitutionally faithful solution would restore Section 101 largely to the role Congress originally gave it.

The Supreme Court decides cases and controversies. It does not rewrite statutes according to changing policy preferences, especially in patent law. Article I, Section 8 gives Congress the power to establish the patent system, and Congress exercised that power through Title 35 of the U.S. Code. Judicially created exceptions therefore present a serious separation-of-powers problem.

My central concern is judicial legislation. Section 101 did not change. The Supreme Court changed its meaning without congressional authorization.

Congress may adopt PERA, pursue another reform, or restore Section 101 to its original function. That choice belongs to Congress.

The needed reform is modest. For most of American history, Section 101 served as a coarse filter that identified broad categories of patentable subject matter. The Patent Act assigned the harder questions elsewhere. Sections 102, 103, and 112 ask whether an invention is novel, nonobvious, and adequately described.

Section 101 need not do all of patent law’s work at once.

That division of labor governed during more than two centuries of technological progress and economic growth. Yet many courts, commentators, and technology users entered the 21st century convinced that the system produced too many patents. They responded by expanding a threshold eligibility test into a broad policy instrument.

That confidence reflects the institutional conceit Hayek warned against. Innovation policy should ask whether legal institutions encourage the discoveries that have yet to occur. That inquiry requires policymakers to account for visible costs and for inventions delayed, abandoned, or never attempted.

First, Do No Harm to Innovation

As senators debate patent eligibility, they should remember the trauma bay in Pittsburgh.

A physician now holds in one hand the product of more than 20 years of scientific research, entrepreneurial judgment, private investment, technology transfer, clinical validation, patent prosecution, manufacturing expertise, regulatory review, and commercial development. No planner designed that journey. No single institution possessed all the knowledge needed to complete it.

The achievement is that the system succeeded. The troubling question is why success required so much time, money, and persistence.

The greatest cost of a weakened innovation system does not appear in investment totals or patent counts. It appears in the inventions that never emerge. The startup is never formed. The investment is never made. The diagnostic never reaches patients. The entrepreneur decides the odds are too poor and walks away.

Those losses cannot be measured because they never become visible. That makes them easy to ignore, but no less real.

The blood test now improving trauma care in Pittsburgh offers more than a medical success story. It shows what free people can accomplish when secure property rights, voluntary exchange, and institutional humility allow dispersed knowledge to become useful innovation.

Congress cannot plan the next breakthrough. It can stop making the path harder.

The post A Patent, a Blood Test, and 20 Years of Waiting appeared first on Truth on the Market.

Artificial intelligence data centers, advanced manufacturing, and cryptocurrency operations are driving electricity demand sharply higher. Yet the rules for connecting new power sources and major users remain slow, fragmented, and rigid. The grid’s emerging bottleneck is therefore less a failure of technology than of institutions.

The Federal Energy Regulatory Commission (FERC) has begun to confront that mismatch. It is pursuing two major reform tracks: a series of Section 206 “show cause” orders focused on connecting large electricity users to the grid, and an overhaul of the Natural Gas Blanket Certificate program aimed at speeding infrastructure approvals.

That turn toward market realities is welcome. But success will depend on whether FERC resists the temptation to replace one thicket of rules with another. Its final policies should instead draw on basic law & economics principles: reduce transaction costs, clarify property rights, and avoid overconfident central planning.

Reform will also require addressing the regulatory friction between FERC and the Nuclear Regulatory Commission (NRC), which can leave even promising projects trapped between agencies.

Let the Regions Compete

In its large-load directive, issued in response to the U.S. Department of Energy, FERC ordered the nation’s six jurisdictional regional grid operators—the PJM Interconnection (PJM), the Midcontinent Independent System Operator (MISO), the Southwest Power Pool (SPP), the California Independent System Operator (CAISO), the New York Independent System Operator (NYISO), and ISO New England (ISO-NE)—to overhaul how they process transmission requests and evaluate co-located generation. The goal is to cut review times to roughly 60 to 90 days while protecting existing ratepayers.

From a market perspective, FERC’s decision to reject a rigid, “one-size-fits-all” national rule is exactly right. As I have argued in other contexts involving rapidly evolving technologies, centralized planning often suppresses the experimentation that drives better outcomes. By using regional “show cause” proceedings instead, FERC allows different market designs to compete in real time.

That process is already underway as grid operators race to meet the 2026 compliance deadlines. FERC’s co-location order directed at PJM, for example, found the absence of formal tariff rules governing behind-the-meter generation (BTMG)—power generated on-site rather than drawn from the grid—to be unjust and unreasonable.

PJM’s stakeholders responded by proposing clearer transmission-service categories that distinguish among Network Integration Transmission Service (NITS), firm contract demand, and non-firm contract demand. The goal is simple: align transmission charges more closely with how much each customer actually relies on the grid.

The proposal also seeks to reduce cost-shifting. It would cap retail BTMG netting at a cumulative 50-megawatt threshold, ensuring that large data centers bear the network costs they create while exempting emergency backup generators needed for reliability. To reduce the growing backlog of interconnection requests, PJM also created an Expedited Interconnection Track for shovel-ready generation projects larger than 250 megawatts, with a 10-month review target backed by substantial, nonrefundable readiness deposits to discourage speculative applications.

Other regions are testing different approaches. SPP, for example, has proposed its High Impact Large Load (HILL) framework and Conditional High Impact Large Load Service (CHILLS), which would give major electricity users expedited, non-firm transmission service for up to seven years while permanent network upgrades are completed. This diversity of regional approaches gives the market an opportunity to discover which contractual arrangements reduce transaction costs most effectively.

As FERC finalizes these reforms, it should pair “user-pays” obligations with strong property rights. If a data center developer finances a major transmission upgrade, it should receive transferable capacity rights or credits. Turning a regulatory obligation into a marketable asset would encourage private investment in grid infrastructure while accelerating grid expansion.

The Courts Pull the Plug

As FERC moves to complete these infrastructure reforms, it must also contend with a more skeptical federal judiciary. Agencies have long treated phrases such as “the public interest” and “public convenience” as elastic grants of authority. Recent appellate decisions suggest that era is ending.

Consider the D.C. Circuit’s decision vacating FERC’s approval of the Transco natural-gas project. The court did not merely send the matter back for further explanation. It struck down the authorization because FERC had failed to substantiate “market need” under Section 7 of the Natural Gas Act. The message was blunt: balancing the public interest is not a license for agency intuition. It requires evidence that a project’s benefits outweigh its costs and market distortions.

The D.C. Circuit made a similar point in Affirmed Energy LLC v. FERC. FERC has broad discretion to accept tariff revisions, but it still must offer a rational, evidence-based explanation. Otherwise, its decision risks being arbitrary and capricious.

The 3rd U.S. Circuit Court of Appeals reinforced the other side of that boundary in Transource Pennsylvania LLC v. DeFrank. The court held that state utility commissions may not independently second-guess a FERC-approved regional plan to reduce transmission congestion. Federal market rules may be constrained, but states cannot casually override them for protectionist ends.

FERC’s most important limit may come from the Supreme Court’s major-questions doctrine. In West Virginia v. EPA (2022), the court held that agencies addressing matters of “vast economic and political significance” must point to clear congressional authorization rather than vague or open-ended statutory language.

That principle matters because advocates of broader federal planning have often urged FERC to transform its duty to ensure “just and reasonable” wholesale rates into a mandate for sweeping environmental policy or government-directed shifts in energy resources. As Commissioner Mark Christie argued in his dissent from the Certificate Policy Statement, using the Natural Gas Act to deny infrastructure permits based on broad policy goals, including economywide greenhouse-gas projections, risks triggering the major-questions doctrine.

Energy policy has enormous economic consequences, but that does not give FERC a roving commission to redesign the grid. Vague appeals to the “public interest” cannot substitute for authority Congress never granted.

The lesson for FERC’s review of the Natural Gas Blanket Certificate Program is straightforward. By roughly doubling the cost thresholds below which pipeline and hydroelectric operators may proceed without project-by-project approval, FERC would reduce the costs and delays of administrative gatekeeping. Case-by-case certification can act as a barrier to entry, constraining supply and raising prices.

Predictable, quantitative rules offer a better path than open-ended public-interest balancing. To capture those gains and survive judicial review, FERC should tie any expansion of blanket authority to clear statutory mandates for infrastructure efficiency and automatically index the new cost thresholds to inflation.

Caught Between Two Regulators

FERC’s current reform agenda is a welcome start. But from a law & economics perspective, it leaves the biggest long-term obstacle largely untouched: the regulatory divide between FERC and the NRC.

The market is already signaling strong demand for next-generation nuclear power to supply energy-intensive facilities such as artificial intelligence data centers. One promising model is to co-locate a data center with a nuclear plant, allowing it to draw electricity directly from the facility “behind the meter” rather than through the broader transmission grid.

That arrangement, though, falls into a regulatory gap. FERC oversees wholesale electricity markets and grid interconnection, while the NRC regulates nuclear safety. Connecting a large data center directly to a nuclear plant can change the facility’s operating profile, affecting everything from reliability planning to emergency procedures.

Because FERC and the NRC lack a unified framework for evaluating these hybrid projects, developers face overlapping reviews, high transaction costs, and delays that can stretch for years.

The problem extends beyond jurisdictional overlap. The NRC has long relied on a highly prescriptive, zero-risk approach that often overlooks the economic costs of keeping reliable, carbon-free generation offline. A more risk-informed, performance-based regulatory framework would focus on measurable safety outcomes while giving developers greater flexibility in how they achieve them.

FERC also has a role to play. Its capacity markets should more accurately value the reliability benefits that nuclear generation provides, including its ability to reduce volatility during periods of high demand. Pricing those benefits directly would allow market incentives—not state intervention—to guide investment in dependable generating resources.

Get Out of the Grid’s Way

FERC’s reforms are a promising step toward easing the regulatory drag on America’s energy infrastructure. Stronger user-pays rules, clearer property rights for investors, and broader blanket-permitting authority could help the grid expand faster and more efficiently.

But the easy fixes will only go so far. Unless FERC and the NRC resolve the regulatory stalemate over co-located nuclear power, the government will keep throttling the energy growth it says it wants.

The grid does not need another plan. It needs regulators to get out of one another’s way.

The post Lightning in a Bottleneck: How Regulatory Gridlock Is Choking America’s Energy Boom appeared first on Truth on the Market.

Broadband providers rarely own the poles their networks use. Reaching a new community usually requires attaching equipment to poles owned by a utility or another company. Who pays the resulting costs can determine whether a rural build makes economic sense—or never gets built.

The trouble often starts with the poles themselves. Many are old, overloaded, or burdened by safety and code violations left by earlier attachers. Pole owners may then try to make the newest provider pay to fix problems it did not create.

Appalachian Power Co.’s (APCo) “cost causer” policy put that practice squarely before the Federal Communications Commission (FCC). When Comcast sought to attach to a pole carrying someone else’s violation, APCo required Comcast to pay to replace the entire pole. In effect, APCo treated the arrival of a broadband builder as an opportunity to shift deferred-maintenance costs onto the party least responsible for them.

In February 2026, the FCC sided with Comcast and struck down the policy. In the first decision under its accelerated docket for pole-attachment disputes, the FCC held that APCo could not force a new attacher to pay the full cost of replacing a pole that was already out of compliance. The new attacher could be charged only for the incremental cost of installing a taller or stronger pole needed to accommodate its equipment beyond what correcting the existing violation would require.

The ruling should have settled the matter. It did not.

Rather than apply the FCC’s cost-sharing standard, APCo continued billing Comcast a flat 20% of the full replacement cost, whether or not Comcast’s attachments caused that expense. Comcast returned to the FCC earlier this month with a new complaint.

The dispute now tests more than one utility’s billing policy. It will show whether the FCC’s landmark broadband-deployment order has real force—or merely looks good on paper.

The Pole Was Broken Before Comcast Got There

Broadband and cable providers cannot build networks without access to the utility poles along their routes, yet they rarely own those poles. Section 224 of the Communications Act therefore requires pole-owning utilities to offer access on “just and reasonable” rates, terms, and conditions. 

The exception is narrow. An electric utility may deny access on a nondiscriminatory basis only when a pole lacks capacity or when safety, reliability, or generally applicable engineering concerns require it. In states such as Virginia, which do not regulate attachments to electric-utility poles, the FCC oversees those rates, terms, and conditions. 

Most pole attachments require some preparatory work. The central economic question is who pays the “make-ready” costs—the labor and materials needed to prepare a pole for new equipment, sometimes including replacement of the pole itself.

Two rules govern that allocation. The general rule, 47 C.F.R. § 1.1408(b), requires parties that gain access to or directly benefit from a modification to share its cost proportionately. A more specific rule, adopted in the FCC’s 2018 Pole Attachment Order and codified at 47 C.F.R. § 1.1411(e)(4), bars utilities from charging a new attacher to correct violations caused by someone else before the new attachment. 

Those protections matter most in unserved and underserved areas, where unnecessary replacement costs can kill a broadband project before construction begins. In Virginia, Comcast must attach to poles owned by APCo, some of which already carry code or construction-standard violations caused by other parties. 

Under the 2025 policy that was challenged before the FCC, APCo required the new attacher to pay 100% of the replacement cost up front, subject to a possible 50% refund. The preexisting violator could not move its equipment to the replacement pole unless it paid the other half.

The catch was obvious. Comcast received a refund only if the violator chose to transfer its equipment and paid its share. If the violator walked away, Comcast recovered nothing. Either way, Comcast remained responsible for between half and all of a replacement cost caused by someone else.

When Pole Costs Pull the Plug

Who pays for replacing a utility pole is not an accounting detail. It can determine whether a rural broadband project gets built at all.

The economics of rural deployment are especially unforgiving. In cities and suburbs, each pole serves many homes, spreading its fixed cost across a large subscriber base. In the unserved and underserved communities targeted by programs such as the Broadband Equity, Access, and Deployment (BEAD) Program and the earlier Rural Digital Opportunity Fund (RDOF), providers often need several poles to reach a single additional home. Every dollar spent on pole replacements is therefore spread across far fewer customers, making those costs disproportionately important where expanding broadband matters most.

Providers’ experience bears this out. Charter Communications, for example, reported that pole-replacement charges accounted for roughly one-quarter of the construction cost for a rural build serving about 57,000 locations.

The problem is not just the size of the bill, but its unpredictability. Providers often do not discover which poles have disqualifying conditions until they survey a route, after they have committed to the project and, for subsidized builds, submitted bids based on much lower expected costs. Because utilities would eventually have to replace aging poles anyway, requiring a new attacher to pay the entire replacement cost charges far more than the costs the attachment actually causes. It simply accelerates an expense the utility would otherwise bear, while injecting substantial uncertainty into projects that already operate on thin margins. 

That uncertainty has real consequences. In April 2024, Charter notified the FCC that it was relinquishing RDOF awards covering locations in three states because unexpected pole-replacement costs had made the projects uneconomical.

When replacement costs are both unbounded and assigned to the new attacher, abandoning a marginal project becomes the rational business decision. The FCC has recognized as much, warning that forcing attachers to bear costs they did not cause diverts scarce capital away from network expansion. APCo’s policy exemplified that problem, transforming uncertain liabilities into certain—and often substantial—costs across deployments involving thousands of poles.

You Break It, You Buy It

The FCC ruled for Comcast in February 2026 and struck down APCo’s policy. The decision closely followed the agency’s rules. Section 1.1411(e)(4) prohibits a utility from charging a new attacher to bring a pole, attachment, or third-party equipment into compliance when the violation resulted from someone else’s earlier work. That is precisely what APCo had done. The poles were already out of compliance before Comcast applied, yet APCo required Comcast to pay the full cost of replacing them. The FCC concluded that this violated both the regulation and the agency’s long-standing principle that a new attacher pays only the costs its own attachment actually causes. 

The FCC also rejected APCo’s central defense: that Comcast was the ultimate “cost causer” because the original violator might someday remove its equipment, leaving Comcast as the only remaining beneficiary of the replacement. The agency found that argument too speculative. APCo presented no evidence that violators were removing their equipment, and its own policy contemplated that the violating attachments could remain on a “stub pole” indefinitely. Nor did APCo’s promise of a possible 50% refund solve the problem. A provider forced to pay the entire replacement cost up front is not made whole by a partial reimbursement that depends on someone else’s future choices. If anyone should bear the remediation costs, the FCC concluded, it is the party that caused the violation. 

The FCC did not, though, embrace Comcast’s broader position that a preexisting violation eliminates any obligation for the new attacher. Instead, it harmonized the two governing rules. The existing violator must pay to correct the preexisting problem. But if the new attachment requires a pole that is taller or stronger than the one needed simply to remedy that violation, the new attacher must pay the incremental cost under Section 1.1408(b).

The FCC illustrated the distinction with a simple example. Suppose fixing the existing violation requires replacing a 40-foot pole with a 45-foot pole costing $5,000. If accommodating the new attacher instead requires a 50-foot pole costing $5,500, the violator pays the $5,000 needed to correct the existing deficiency, while the new attacher pays only the additional $500. In short, the FCC drew a straightforward line: correcting someone else’s violation is not the newcomer’s responsibility, but paying for the extra capacity its own attachment requires is.

Twenty Percent of Wrong

A clear rule matters only if utilities follow it. Here, implementation has become the problem.

Rather than charge Comcast the incremental cost the FCC described, APCo has continued billing a share of the entire replacement cost—a blanket minimum of roughly 20% per pole. As Comcast argues, that figure has no meaningful connection to the FCC’s standard. The extra cost of installing a taller or stronger pole may be only $100 to $200. By contrast, 20% of a full replacement costing several hundred or several thousand dollars sweeps back in the very remediation expenses the FCC said Comcast cannot be forced to pay.  

APCo has also suggested that 20% is a floor, not a ceiling, and that it may seek an even larger share. That turns the FCC’s holding on its head. The agency did not invite APCo to devise whatever split it considered fair. It adopted a specific cost-causation rule: Comcast must pay only for the additional pole capacity its attachment requires. 

A flat percentage of the total replacement cost is not that. It is the same cost shift the FCC rejected in February, dressed up as compliance. Multiplied across thousands of poles and pressed against federal funding deadlines, even a modest-looking percentage becomes exactly the kind of unpredictable, deployment-deterring expense the rules were designed to prevent. 

The Bill Comes Due

Stripped of its technical details, APCo’s policy required a broadband provider to pay for a problem it did not create, on poles someone else left out of compliance, in communities that federal and state programs are spending billions to connect.

The FCC rejected that arrangement and drew a sensible line. The party responsible for a preexisting violation must pay to correct it. A new attacher must pay only for the additional pole capacity its own equipment requires. That rule vindicated Comcast’s position and, more importantly, protected the broadband-deployment goals behind the agency’s regulations.

What remains is enforcement. By continuing to charge a fixed share of full replacement costs despite the FCC’s order, APCo is testing whether the agency’s first accelerated-docket pole-attachment ruling binds utilities in practice or only on paper.

The FCC has already said who should pay. It should not let APCo send the bill elsewhere.

The post The Cost of Holding Up Broadband appeared first on Truth on the Market.

For the past two years, warnings about life insurers’ private-credit investments have become a staple of financial commentary. In 2024, the International Monetary Fund cautioned that private credit’s rapid growth and limited transparency could eventually threaten the broader financial system. Researchers at the Federal Reserve Bank of Boston examined whether the sector’s expansion poses stability risks, especially through bank credit lines to private-credit funds. Moody’s has flagged liquidity and concentration risks in insurers’ growing allocations. Axios suggested that annuities could transmit private-credit losses to ordinary households, while a Forbes columnist argued that rising defaults are already testing both banks and insurers.

The concern is not hard to understand. Life insurers hold $9.9 trillion in general-account assets against annuity and life-insurance obligations that may stretch decades into the future. Private loans do not trade on public markets, so insurers value them through models and appraisals rather than observable prices. Many also carry private ratings disclosed to the insurer but not to the market. If those valuations or ratings prove too rosy, insurers may hold less capital than their actual risks require—and the gap may surface only when policyholders expect payment.

That makes private credit a plausible source of risk. It does not make it a proven one. Whether the danger exists in practice is an empirical question, and regulators already collect the data needed to answer it.

In a recent International Center for Law & Economics (ICLE) white paper, Lars Powell and I used National Association of Insurance Commissioners (NAIC) annual-statement data to test whether life insurers with larger private-credit holdings are financially weaker than their peers. They are not.

Private Credit: Define Your Terms

The first problem is definitional: “private credit” has no settled meaning. The term generally covers lending outside public bond markets and traditional bank balance sheets, especially direct loans to midsized companies and structured securities backed by cash flows such as equipment leases or receivables. But insurers’ regulatory filings contain no line item labeled “private credit,” so researchers must decide what counts.

That choice matters. Recent estimates of insurers’ exposure range from $289 billion to $1.8 trillion, while a Federal Reserve Bank of Chicago working paper puts the figure at $849 billion. The highest estimates rely on definitions broad enough to include privately placed bonds that differ little from public bonds in liquidity or transparency.

Our study uses a deliberately narrow definition aimed at the assets critics actually worry about: privately placed debt classified as direct loans, plus nonmortgage structured securities with private-letter ratings. On that basis, private credit accounted for about 6% of life insurers’ general-account assets in 2025, up from roughly 3% in 2020. Annual growth has also slowed sharply as insurers approach their target allocations.

The holdings are unevenly distributed. Of 342 insurer groups, more than half report no private-credit exposure at all. Among those that do, the mean allocation is 4.2% of assets, the median is 3.1%, and the 95th percentile is 10.8%.

Those proportions offer some reassurance. Whatever risk private credit poses must depend partly on how much of an insurer’s portfolio it occupies. An insurer with half its assets in hard-to-value loans would merit close scrutiny. One with 3% could likely absorb even a severe write-down through surplus. Nearly every U.S. life insurer looks much more like the latter, and diversification rules and concentration limits help keep it that way.

The Data Decline to Panic

Small allocations could still matter if private credit were weakening insurers’ financial health. To test that possibility, we estimated the relationship between private-credit holdings and the risk that a life insurer becomes insolvent.

Using NAIC annual-statement data and the historical record of insurer failures, we estimated the probability that each insurer would enter formal regulatory proceedings within the next one or two years. The model incorporated financial ratios, as well as measures of size, leverage, profitability, organizational structure, and group affiliation. We then tested whether those estimated insolvency probabilities were related to the share of assets invested in private credit through two regression analyses.

The first regression controlled for economywide and industrywide conditions affecting all insurers. It found a negative relationship: insurers with larger private-credit allocations had lower estimated insolvency risk on average. The result was statistically significant at the 1% level, although private-credit exposure explained only a small share of the overall variation in financial strength.

A natural objection is that financially stronger insurers may simply be more likely to invest in private credit. To address that possibility, we compared each insurer against its own history. If private credit weakens insurers, then increasing an insurer’s allocation should raise its insolvency risk. It did not. Changes in an insurer’s private-credit holdings showed no statistically detectable relationship with its estimated probability of insolvency.

None of this proves that private credit makes insurers safer. It does show that insurers with larger private-credit portfolios are not financially weaker than their peers and, on average, appear somewhat stronger. It also shows that, during the sample period, increasing private-credit exposure within an insurer was not associated with greater insolvency risk.

Two caveats are worth keeping in mind. First, these are correlations, not proof that private credit causes stronger financial performance. Second, the findings reflect insurers’ current portfolios, where private credit rarely exceeds 11% of assets. They do not speak to portfolios far outside that range. Within the range that actually matters for current policy, though, the claim that private credit is undermining insurer solvency finds no support in the data.

A Match Made in Maturity

The economics also point in the same direction. Start with a simple reality: life insurers promise to make payments decades into the future. They therefore need assets that generate predictable cash flows over similarly long horizons. Economists call this asset-liability matching, and long-term private loans are well suited to the job.

Private credit also tends to earn higher returns for a reason. Nobel laureate Oliver Williamson described assets like these as “idiosyncratic.” Unlike publicly traded bonds, each loan is individually negotiated, subject to its own covenants, and typically held until maturity. That makes the loans costly to evaluate and difficult to sell. Investors who might need to exit quickly demand extra compensation for accepting that illiquidity. Researchers at the Federal Reserve Bank of Chicago estimate the premium at as much as 80 basis points over comparable public bonds.

Life insurers are in a different position. Because they generally expect to hold these loans to maturity, they can collect the illiquidity premium without bearing much of the liquidity risk that justifies it in the first place.

Private credit may also benefit the broader economy. Analyzing more than 18,000 loans, Franz Hinzen and his co-authors find that when banks tighten lending standards, borrowers shift to private-credit funds in large numbers. That substitution helps keep credit flowing to businesses, supporting investment and employment during downturns.

Private Credit Is Not a Bank

Much of the concern about private credit comes from applying risk models designed for banks. That is a category error. Bank risk and private-credit risk are fundamentally different.

Banks fund long-term loans with demand deposits that customers can withdraw at any time. If confidence falters, depositors may rush to pull their money, forcing banks to sell assets at distressed prices and potentially triggering a bank run.

Private-credit funds operate very differently. In the largest study of private-credit fund balance sheets to date, covering roughly 1,300 funds and 60% to 70% of U.S. private-credit assets, Gregor Matvos, Tomasz Piskorski, and Amit Seru found that the median fund finances 98% of its assets with investor equity committed for at least a decade. By comparison, equity accounts for only about 11% of U.S. commercial banks’ funding.

That difference matters. Private-credit funds typically last 10 to 12 years, longer than the loans they hold, so investors cannot demand their money back before the assets mature. When losses occur, they are absorbed by institutional investors who knowingly accepted that risk.

Business development companies (BDCs)—the primary vehicles for direct lending—are also conservatively financed. Federal law limits their borrowing, and they averaged about $0.91 of debt for every dollar of equity in early 2025. Sergey Chernenko, Robert Ialenti, and David Scharfstein estimate that the median BDC holds capital equal to roughly 36% of risk-weighted assets, compared with about 13% for the large banks in the Federal Reserve’s stress tests.

Reviewing this evidence in its 2025 annual report to Congress, the Office of Financial Research concluded that private credit poses limited financial-stability risk because the sector relies on low leverage and long-term funding. Where meaningful links to the banking system do exist—notably banks’ credit lines to private-credit funds—regulators already receive detailed, loan-level reporting from the largest banks.

Oversight Is Not the Missing Piece

Some concern about insurers’ private-credit holdings rests on a mistaken premise: that the relevant firms operate in a regulatory void. They do not. Private lenders, private-credit funds, and life insurers all face oversight, though through different regimes.

BDCs register under the Investment Company Act, file annual and quarterly reports with full investment schedules, and must value illiquid assets in good faith under Securities and Exchange Commission (SEC) rules. Advisers to private-credit funds operate under the Investment Advisers Act, owe fiduciary duties to their clients, and submit confidential SEC reports on holdings, borrowing, liquidity, and performance. Banks’ loans to the sector also appear in regulatory filings.

Life insurers, meanwhile, operate under state solvency regimes that impose risk-based capital requirements, conservative statutory accounting, diversification rules, concentration limits, own-risk assessments, periodic examinations, and continuous supervision.

The real policy question, then, is not whether regulators are watching. It is whether private credit creates risks that this combined system fails to detect. So far, the evidence offers little reason to think it does.

That point matters because the same state-based insurance regime carried the life-insurance industry through the 2008 financial crisis. And the assets that caused the damage then were not obscure private loans. They were publicly traded securities bearing published ratings from the largest credit-rating agencies.

Regulate the Risk, Not the Headline

Private credit raises legitimate questions about valuation, data quality, bank-nonbank connections, private-equity ownership of insurers, and offshore reinsurance. Those issues deserve close regulatory attention. They do not, by themselves, show that private credit is weakening insurer solvency.

That distinction matters. The evidence points toward better disclosure, stronger data collection, closer supervisory coordination, and more refined measures of risk—not broad restrictions or punitive capital charges.

Heavy-handed rules could backfire. They would reduce insurers’ access to assets that are well matched to their long-term liabilities while pushing more lending into less-transparent markets. That would leave regulators with less visibility and the financial system no safer.

The Perils of Assuming the Worst

Regulators are not standing still. The NAIC is already reviewing how insurers’ private-credit holdings should be treated for capital purposes.

A bit of background helps. Every security an insurer owns receives a risk designation that determines how much capital the insurer must hold against it. For most bonds, ratings from recognized credit-rating agencies automatically translate into those designations through a process known as “filing exemption.” Many private placements instead rely on private-letter ratings, which are provided only to the insurer. In 2024, more than 152,000 securities received regulatory treatment through one of those two mechanisms.

The NAIC is now reassessing how filing-exempt and private-letter ratings should translate into capital requirements. Its basic approach is sound. Rather than second-guessing individual ratings, it evaluates each rating provider’s overall performance.

The details, though, matter. The current proposal would treat asset classes with limited data or short performance histories as high risk by default. That risks conflating uncertainty with danger. Private assets naturally generate less public data than exchange-traded securities, but that is not the same as proving they are riskier. When evidence is limited, the better response is to collect more of it, not to assume the worst.

The proposal also includes powerful remedies, including withdrawing filing-exempt status from an entire asset class. Those tools could disrupt insurers’ portfolios and the broader credit market. They should be reserved for cases with clear evidence of persistent rating failures, applied prospectively, and paired with lengthy transition periods.

The NAIC is also revising risk-based capital charges for collateralized loan obligations (CLOs), which are securities backed by pools of corporate loans, and plans to extend that work to other asset-backed securities. The proposed charges rely on modeling assumptions that are substantially more conservative than those applied to ordinary corporate bonds, despite CLOs’ strong historical performance, including through the 2008 financial crisis. In comments to the NAIC, ICLE recommended a different sequence: improve data and disclosure first, revise capital factors only as the evidence warrants, and avoid sharp regulatory thresholds that invite arbitrage.

Europe offers a cautionary tale. Solvency II imposed capital charges on securitized assets that market participants—and eventually the European Commission itself—concluded overstated their actual risk. European insurers largely exited the market until the Commission moderated those requirements. U.S. regulators need not repeat that mistake.

Don’t Mistake Private for Peril

Reforms should be prospective, transparent, and tied to evidence of actual risk. Sweeping restrictions or punitive capital charges would impose costs on both sides of the balance sheet. Insurers would lose access to long-duration assets that fit their liabilities unusually well, while lending would migrate to corners of the financial system with fewer disclosures and less supervision. Regulators would see less, not more.

Policyholders would pay, too. The extra yield on private credit helps insurers offer more competitive annuity and life-insurance products. Overcorrection would raise costs without any demonstrated gain in solvency.

Private credit deserves continued scrutiny, and the current supervisory attention is appropriate. But the case that it threatens the life-insurance industry remains unproven, while the available evidence points the other way. Regulators should improve the system through better data, disclosure, and risk measurement—not punish an asset class for the sin of being private.

The post Private Credit, Public Panic: Why Life Insurers Are Stronger Than the Headlines Suggest appeared first on Truth on the Market.

The CJEU’s final judgment in Google Android, handed down last week, will be celebrated in Brussels as a triumph of public enforcement over Big Tech. It deserves a less triumphant reading.

The judgment ends an eight-year legal fight by leaving Google and Alphabet with a fine of roughly €4.125 billion for contractual practices tied to Android, Google Search, Chrome, and the Play Store. The court accepted that Google abused its dominance by using Android distribution terms, preinstallation conditions, anti-fragmentation obligations, and related arrangements to favor its own search and browser products.

The fine is painful. The precedent is worse.

The CJEU approved important parts of the General Court’s analysis. It allowed courts to consider economic context without systematically constructing a counterfactual. It also confirmed that liability does not always depend on proof that the practices could foreclose an “as-efficient competitor” (AEC). That test asks whether a rival as efficient as the dominant firm could compete despite the challenged conduct.

That doctrinal signal matters more than the penalty. A €4 billion fine stings. But the larger cost comes from the precedent’s effects on platform design, investment incentives, and the legal expectations facing future European technology firms. If Article 102 of the Treaty on the Functioning of the European Union (TFEU)—which governs abuse of dominance—condemns ordinary platform governance whenever rivals dislike the outcome, Europe will not get more innovation. It will get more litigation, more regulatory redesign of products, and fewer firms willing to build integrated platforms in the first place.

The Trouble With Treating Scale as Suspicious

This is an odd place for Europe to be. Mario Draghi’s report on European competitiveness warned that Europe largely missed the internet-led digital revolution. It also found that the productivity gap between the European Union and the United States came largely from the technology sector, and that only four of the world’s top 50 technology companies are European.

When it comes to closing the innovation gap with the United States and China, the numbers are grim. No EU company with a market capitalization above €100 billion has been built from scratch in the past 50 years. European firms spent €270 billion less on research and innovation than their U.S. counterparts in 2021. Nearly 30% of European-founded unicorns relocated abroad between 2008 and 2021, overwhelmingly to the United States.

Alas, the Google Android approach points in precisely the wrong direction. Europe needs more risk-taking, not more legal suspicion of business success. It needs more scale, not more hostility to the contractual tools that help platforms coordinate device makers, app developers, consumers, advertisers, and service providers. It needs more experimentation in monetization, default design, distribution, and integration, not more efforts to recast successful platform architecture as anticompetitive conduct after the fact.

Android is not a conventional product sold in isolation. It is a mobile operating system that links handset makers, app developers, carriers, consumers, advertisers, browser developers, search providers, and Google itself. Google made Android broadly available, limited fragmentation, supported developers, maintained compatibility, and monetized much of the project through complementary services such as search.

That model can be criticized, but it is not obviously anticompetitive. Even if it were, the central law & economics question remains: compared with what?

Compared with a world in which Google could not assure prominent distribution for Search and Chrome, would Android have been offered as broadly, cheaply, and consistently? Would handset makers have received the same free or low-cost operating system? Would developers have enjoyed the same compatibility and reach? Would consumers have received lower prices, better app availability, safer devices, and stronger competition against Apple’s vertically integrated iOS model?

Those questions are not side issues. They are the core of the case. A platform’s defaults, bundles, and compatibility rules often are not dirty tricks. They are part of the machinery that allows the platform to exist.

International Center for Law & Economics (ICLE) scholar Dirk Auer made this point years ago. His 2020 analysis of the European Commission’s Android decision argued that the Commission’s factual account did not prove that Google harmed competition or consumers. He criticized the decision’s market definition, dominance analysis, treatment of Apple, theories of harm, and innovation claims.

Two years earlier, Auer warned that the Android decision threatened the viability of Android’s open-source model by meddling with the governance rules through which Google competed with Apple. One need not accept every detail of those critiques to see that the core warning has aged well. Competition law becomes dangerous when it treats platform coordination as suspect merely because rivals would prefer different terms.

Geoffrey Manne, Lazar Radic, Dirk Auer, and other ICLE authors have made the same broader point. Digital competition rules often dress up competitor protection as competition policy. Their critique is not that digital platforms can do no wrong. It is that many recent rules and decisions prioritize fairness, contestability, and redistribution among firms over efficiency, innovation, and consumer welfare.

That approach can shield less efficient rivals, dilute the advantages successful platforms earned, and make firms less willing to invest in the next platform. Google Android shows the problem. A rival search engine may want guaranteed placement. A rival browser may want default status. A device maker may want Google’s app store without Google’s compatibility rules. Those are understandable business preferences. They are not evidence that consumers were harmed.

Defaults Are Design, Not a Smoking Gun

The Android judgment is part of a larger project. It belongs with Google Shopping (2024) and Europe’s broader campaign against self-preferencing, platform integration, and so-called gatekeeper advantages.

In Google Shopping, the CJEU upheld an expansive theory of leveraging through the design and display of search results. The court also confirmed that the Commission did not need to reconstruct a full counterfactual or show foreclosure of an as-efficient competitor in the way a more disciplined effects-based inquiry might require.

The common thread is not a traditional showing that consumers paid more, output fell, or innovation slowed. It is the intuition that a successful platform’s control over valuable distribution is suspect when rivals would prefer another allocation. That intuition shifts antitrust away from consumer welfare and toward platform-neutrality regulation. It treats search results, defaults, app-store licenses, and product design as quasi-public utilities to be allocated by administrative judgment.

High-technology competition rarely works that way. Firms invest in platforms partly because integration can yield advantages. If those advantages later become duties to assist rivals, the incentive to build the next platform weakens before anyone writes the first line of code.

The old antitrust maxim remains indispensable. Competition law should protect competition, not competitors. In digital markets, that maxim must account for dynamic competition. If a firm invests in a platform, attracts developers, subsidizes users, solves compatibility problems, and offers a product consumers value, rivals will often lose ground. That is not a pathology. It is competition among business models. Some are integrated, some are modular, some are ad-funded, and some are subscription-funded.

Defaults illustrate the problem. The Android decision gives great weight to status quo bias, meaning the tendency of consumers to stick with what comes preinstalled. That tendency is real, but it does not prove anticompetitive harm.

Defaults reduce search costs. They make a device usable out of the box. They can signal quality. They can fund a zero-price operating system. Users can change them. Rivals can counter them through brand, distribution, and product quality. To say that a default matters is not to say that it excludes competition. The relevant question is whether the default raises quality-adjusted prices, reduces output, degrades quality, suppresses innovation, or prevents equally efficient rivals from reaching consumers through realistic channels.

A consumer-welfare-oriented court would insist on a rigorous counterfactual. Without one, enforcement can confuse the observed success of an integrated strategy with proof that the strategy caused unlawful foreclosure.

That matters even more in platform markets, because an intervention on one side can harm participants on another. Requiring Google to unbundle, weaken defaults, or tolerate fragmentation may help rival search engines or browsers. It may also reduce Google’s incentive to invest in Android, increase device makers’ costs, reduce security, complicate developer support, or make Android less attractive against Apple. A serious counterfactual would compare those possibilities. The Android judgment’s tolerance for avoiding that analysis invites false positives.

The court’s treatment of the as-efficient-competitor principle raises the same concern. The AEC test is not a talisman, and it will not fit every nonprice case. But the principle behind it is vital. Article 102 should not preserve rivals who lose because they are less attractive to consumers, less innovative, less efficient, or slower to develop alternative distribution channels. It should intervene when dominant-firm conduct prevents competition on the merits from working.

A doctrine that loosens the AEC inquiry because the market is digital risks turning the special features of technology markets into an excuse for weaker proof.

Article 102 Needs a Consumer-Welfare Spine

The Android judgment departs from better instincts in recent European case law.

In Intel (2017), the CJEU required careful attention to evidence and economic analysis when a dominant firm argued that its rebates could not foreclose as-efficient competitors. In Servizio Elettrico Nazionale (2022), the court tied Article 102 to practices that may harm consumers, including indirectly through harm to the competitive process. It also recognized that a dominant firm may justify its conduct by showing positive consumer effects. In Unilever Italia (2023), the court held that exclusivity clauses must be capable of exclusionary effects and that authorities must assess evidence submitted by the dominant undertaking, including economic studies and AEC-type evidence where relevant.

Those decisions do not create a U.S.-style antitrust code. Europe need not copy American doctrine wholesale. But they display virtues European competition policy badly needs: limiting principles, attention to effects, seriousness about evidence, and a willingness to distinguish exclusion caused by anticompetitive conduct from exclusion caused by superior performance.

ICLE’s 2024 comments on the Commission’s draft Article 102 guidelines describe this more effects-based line of cases as including Intel, Servizio Elettrico Nazionale, Unilever Italia, and Intel Renvoi. The comments also criticize the Commission’s more formalistic turn, which risks chilling procompetitive conduct when Europe already trails in productivity and competitiveness.

In 2024, I argued that Europe’s latest monopolization-policy direction threatens innovative business practices that promote high-tech growth. I criticized presumptions against conduct such as self-preferencing, as Brian Albrecht and Manne also discussed in 2025, along with exclusivity, tying, low pricing, and other practices that may benefit consumers.

The institutional problem is simple. Dominant firms facing vague standards, hard-to-prove efficiency defenses, and massive fines will pull their punches. They will avoid aggressive product improvement, integration, and efficient contracting because the penalty for guessing wrong is enormous.

That is the real danger of Android. The judgment tells large platforms, and aspiring European platforms, that product integration and contractual coordination may later be judged through the lens of rival disadvantage. It tells firms that a successful default can become evidence of status quo bias, monetization tied to distribution can become unlawful leveraging, and compatibility obligations can become obstruction of alternatives. It tells entrepreneurs that the reward for scale may be a decade of litigation and a court-approved redesign of the business model.

Google should not receive immunity because it innovated. Dominance does not create a license to deceive, coerce, sabotage interoperability, or sacrifice product quality merely to block rivals. Nor is every default, bundle, or exclusive arrangement benign. Some practices may exclude equally efficient rivals and harm consumers. But the legal system must prove that harm rather than infer it from the fact that rivals lost a preferred path to users.

The distinction matters because innovation is path-dependent. Firms do not merely respond to today’s fine. They respond to the legal environment they expect over the next decade. If that environment tells them that successful platform governance will be second-guessed, rational firms will design less ambitiously. They will avoid business models that require cross-subsidy. They will offer fewer integrated features. They will litigate before launching. They will negotiate with regulators before testing products with consumers.

None of that appears in a static foreclosure chart. All of it matters for dynamic competition.

The predictable beneficiaries are not necessarily consumers. They are often business users, rivals, and follow-on plaintiffs who can turn a public judgment into private damages claims. Once a court validates a broad theory of exclusion, litigation incentives change. Rivals can seek compensation for lost distribution, not because consumers paid more or received worse products, but because the rival’s path to scale was harder. Competition law then becomes a tool for redistributing returns from platform investment. It stops disciplining harm to competition and starts refereeing disappointment.

The Next Platform Will Read the Fine Print

A wiser Article 102 framework for platform cases would begin with consumer welfare and dynamic competition, not an abstract preference for rival access.

First, courts should require a serious counterfactual. That inquiry should consider the platform’s monetization model, investment incentives, product quality, security, compatibility, and competition among platforms.

Second, when a dominant firm submits plausible evidence that as-efficient rivals could compete, or that the practice creates efficiencies, authorities should engage that evidence rather than dismiss it as unnecessary.

Third, integration, tying, preinstallation, anti-fragmentation rules, and self-preferencing should not be treated as suspect labels. They are business practices. They may help consumers or harm them depending on context.

Fourth, courts should distinguish access to a rival’s created asset from suppression of competition itself. Google’s rivals may want access to Android distribution on terms that maximize their reach. But Android’s distribution architecture was not a natural resource. Google built and maintained it through investment, governance, licensing, developer support, and monetization. Antitrust should hesitate before converting such assets into regulated opportunities for rivals. Forcing a platform to subsidize competitors may sound like fairness. It can also reduce the reward for creating the next platform.

European courts should also remember Draghi’s warning. Europe is not overrun with homegrown global technology platforms. It is struggling to create them. A competition policy that treats scale, integration, default design, and platform control as presumptively suspicious will not close that gap. It will widen it. Europe cannot regulate its way into technological leadership by making the business models of technological leadership legally precarious.

This most recent Android judgment is a setback because the court’s reasoning risks pulling European abuse-of-dominance law away from the economically grounded path suggested by Intel, Servizio Elettrico Nazionale, and Unilever Italia. Future CJEU decisions should return to that path. They should require evidence, counterfactuals, efficiency analysis, and a clear showing of harm to consumers and dynamic competition.

Above all, they should reject the easy slide from harm to businesses to harm to competition. Europe’s greatest high-tech competition problem is not that successful platforms innovate too much. It is that too few firms believe Europe will reward them if they build the next one.

The post Android and the Art of Regulatory Self-Harm appeared first on Truth on the Market.

Sen. Bernie Sanders’ (I-Vt.) proposal would require covered AI developers to transfer up to 50% of their corporate value to a new federal sovereign wealth fund. That fund would distribute “dividends” to the public and use its ownership stake to steer AI development “in the public interest.”

The bill therefore raises questions that go well beyond artificial intelligence. It implicates basic principles of value creation, property rights, corporate governance, political choice, and capital formation. Its central premise is that public data gives rise to public ownership. That premise confuses the availability of information with the entrepreneurial and technical process required to transform information into a productive asset.

This piece argues that the Sanders bill rests on four related errors. First, it treats raw data as the source of economic value, while discounting the entrepreneurial discovery and technical judgment that make data useful. Second, it assumes a federal commission can identify and impose a coherent public interest on a technology marked by conflicting preferences and rapid change. Third, it would weaken the market for corporate control and disrupt integrated firm structures that often reduce transaction costs and improve coordination. Fourth, it would distort capital formation by creating confiscation risk and encouraging firms to organize around a political threshold rather than consumer demand.

Public Data and Private Production

The proposal presents itself as a mechanism for fairness and shared progress. Its core mistake is more basic. It confuses raw information with the creative act of economic production. In doing so, it tries to build a state entitlement on a long-discredited labor theory of value, under which value comes from the inputs people contribute rather than from the entrepreneurial judgment that turns those inputs into something useful.

The problem begins with the Sanders bill’s central claim. Because AI models “ingest” publicly available text, images, and code, the public supposedly has a pre-existing ownership claim to the output. That is the economic equivalent of saying that, because a sculptor uses stone from a public quarry, the surrounding community owns half the statue.

Raw data does not become valuable by existing. Most of it is a disorderly mass of text, images, code, signals, errors, jokes, spam, and half-finished thoughts. It becomes useful only when someone finds a way to structure, filter, and recombine it into a functioning system. Economic historian Joel Mokyr calls this “useful knowledge.” In the AI context, useful knowledge does not reside in the ambient data. It comes from the architecture, training methods, engineering choices, capital investment, and commercial judgment that make the data productive.

Israel Kirzner’s work on distributive justice helps explain why this distinction matters. Kirzner argues that economic value does not emerge through the mechanical conversion of inputs into outputs. It often comes through “entrepreneurial alertness,” or the ability to notice an opportunity that others have missed. Discovery, on this view, is not a planned recipe in which inputs predictably produce outputs. It is an act of seeing and acting. “To notice an opportunity worth grasping is to have created something,” Kirzner writes, and “only she (or he) who has noticed the opportunity and has grasped it, and no one else, is responsible for and is to be credited with the discovery.”

The raw web pages, scientific papers, and social media posts that populate the internet were not “embryonic AI models” waiting for a federal incubator. They were scattered fragments of information. The value of an AI model comes from the speculative and technically demanding work needed to make that information functional. By claiming that the public deserves a 50% equity stake because the raw data was generated by “humanity,” the Sanders bill credits a passive public with an entrepreneurial act it did not perform.

Robert Nozick’s critique of “patterned” distribution in “Anarchy, State, and Utopia” exposes the same error from a different angle. Nozick rejected the idea that there is a neutral, unowned social pot of wealth waiting for planners to divide according to some preferred formula. Wealth does not fall from the sky. It enters the world tied to the labor, capital, risk, and judgment of the people who created it. Production and distribution cannot be cleanly separated.

Nozick argued that political planners often rely on “current time-slice” principles of justice. They take a snapshot of market wealth at one moment and demand that it be rearranged to match a preferred pattern, such as equal shares or state ownership. Maintaining that pattern requires constant state intervention in voluntary exchange. As Nozick put it, the state must “interfere with individual choices.”

Forcing an innovative firm to surrender half its equity to a state-controlled fund treats the creators of that technology as resources for the state to harvest. Nozick described this kind of appropriation as a form of forced labor. The state seizes the fruits of rational planning, speculative risk, and creative effort to fund political ends chosen after the fact.

The bill also collides with the legal history of American intellectual property. Adam Mossoff’s work shows why the bill’s “public resource” rhetoric would overturn a long American understanding of patents, copyrights, and other rights in creative production. Sanders’ bill relies on the progressive legal theory that intellectual property is not real property, but a “public right.” On that view, intellectual property is a regulatory privilege granted by the state, which lawmakers may alter, tax, or reclaim whenever they choose.

Mossoff’s research into the American founding and early patent law refutes that account. The Framers of the U.S. Constitution and early American jurists often treated patents and copyrights as property rights rooted in Lockean principles of self-ownership and labor, not as revocable regulatory favors. John Locke argued in “Two Treatises of Government” that each person has property in his own person, and that the labor of his body and the work of his hands are properly his.

The same principle applies when an inventor or software developer uses the mind rather than the hand. When developers organize information into a working AI system, they create something private through judgment, labor, and risk. Reclassifying that creation as a “public resource” would convert a private right into a state-managed commons. That move would invite the calculational chaos and capital flight discussed in the sections that follow.

Democracy by Committee, Capture by Design

To govern this seized technology empire, the Sanders bill proposes a deceptively simple device: a seven-member “Independent Commission for Democratic AI.” This body would manage the sovereign wealth fund’s massive, non-dilutable equity stake and use its voting power to steer AI companies “in the public interest.”

That sounds tidy enough, which is usually the first warning sign. Once viewed through the Virginia School of Political Economy, which applies economic analysis to political decision-making, the bill’s “democratic” premise collapses into public-choice contradictions and practical impossibilities.

The first problem comes from Kenneth Arrow’s “Social Choice and Individual Values.” Arrow’s Impossibility Theorem shows that, when voters face three or more distinct options, no voting rule can reliably convert their individual rankings into a single, consistent “social welfare function” without violating basic democratic conditions, such as non-dictatorship and transitivity. Transitivity means that, if society prefers A to B and B to C, it should also prefer A to C. Real political preferences often refuse to behave so politely.

AI policy offers a clean example. One faction may want rapid technological deployment to accelerate medical research, scientific discovery, or productivity gains, even if the transition disrupts some jobs. Another may favor strict safety protocols, precautionary pauses, and risk avoidance. A third may want AI development structured to preserve existing jobs, even when that raises costs and reduces productivity.

Those preferences do not point to a single, coherent “public interest” for the commission to pursue. They produce trade-offs that different citizens rank differently. As James Buchanan argued in his critiques of social-choice theory, forcing a single, all-or-nothing “social choice” on a diverse public tends to empower whoever controls the agenda. The seven commissioners would not execute the “will of the people.” They would impose their own politically insulated judgments on a large part of the computational economy, while calling the result “social welfare.”

The deeper irony is that the Sanders bill seeks to “democratize” artificial intelligence by replacing consumer choice with centralized political control. A more direct and flexible democratic mechanism already exists in markets. Markets are imperfect, sometimes badly so, but they allow people to register preferences continuously through prices, purchases, subscriptions, cancellations, and switching.

Ludwig von Mises described this process as consumer sovereignty. In a market economy, consumers direct production by choosing what to buy, how much to buy, and what quality they will pay for. Political voting bundles many issues into one all-or-nothing choice. The pricing system allows more variation. It lets different groups pursue different preferences at the same time.

If some consumers want AI models that emphasize privacy and strict content controls, firms have incentives to serve them. If others want speed, flexibility, or fewer restrictions, other firms can compete for that demand. A market can support multiple models, price points, and risk tolerances. The Sanders bill would replace that process with a commission empowered to impose one politicized standard of “safe and ethical AI” on developers and users alike.

Public-choice theory also warns against the comforting fiction of the benevolent expert. Buchanan and Gordon Tullock’s “The Calculus of Consent” argues that politicians and bureaucrats do not shed self-interest when they enter government. They still respond to incentives, pressure, ambition, reputation, ideology, and organized political demands.

That matters because the Sanders commission would include representatives from specific politically organized groups, including labor unions and safety advocates. A body built around organized interests invites bargaining among those interests. Buchanan and Tullock’s analysis predicts logrolling, rent seeking, and factional bargaining. Commissioners would have strong incentives to trade support across issues, protect their constituencies, and convert long-term technological policy into short-term political payments.

The capture problem is worse. Gordon Tullock’s work on rent seeking shows that, when the state can grant or deny valuable economic privileges, firms shift resources away from production and toward political influence. Money that could fund engineers, chips, data centers, or safety testing instead funds lobbyists, lawyers, and compliance departments. The political machine does not run on fairy dust. It runs on billable hours.

A commission with voting control over up to 50% of covered AI firms would become one of the most valuable targets in Washington. Large incumbents would have the strongest incentives and best resources to shape its decisions. They already have regulatory teams, political relationships, and compliance infrastructure that smaller firms lack.

Those incumbents could press the commission to define “ethical AI safety” in ways that match their existing systems, business models, and compliance capacity. Smaller firms, open-source developers, and disruptive entrants would face higher costs and slower approvals. The result would not be democratic control of artificial intelligence. It would be a state-backed cartel dressed up as public oversight.

The Market for Corporate Control Meets the Ministry of AI

Beyond its errors about knowledge and democracy, Sanders’ AI Sovereign Wealth Fund Act would also damage corporate governance and economic calculation. The bill would intervene in two ways. First, it would take up to a 50% equity and voting stake in covered firms. Second, it would force multi-division technology companies to separate foundational AI research from their non-AI commercial businesses.

That may sound like tidy administrative housekeeping. It is anything but. The bill misunderstands what corporate ownership does and why integrated firms exist. In the process, it would weaken managerial discipline, raise transaction costs, and scramble the price signals that guide investment in high-technology markets.

Henry G. Manne’s corporate-governance scholarship explains why a 50% government voting block would damage corporate efficiency. Manne showed that corporate voting shares are not political ballots for social planning. They are financial instruments that help price control over the firm and discipline managers.

The key mechanism is what Manne called the “market for corporate control.” When managers waste resources, miss opportunities, or pursue goals that reduce firm value, the company’s stock price falls. That lower price can attract outside investors who believe they can run the company better. They may launch a hostile takeover, buy undervalued shares, replace management, and restructure the firm. Even the threat of such a takeover pressures managers to control costs, serve consumers, and allocate capital carefully.

The Sanders bill would short-circuit that mechanism. A non-dilutable 50% government voting block would make hostile takeovers and proxy fights practically impossible at covered firms. No outside investor could replace management over the objection of a permanent federal shareholder with veto power.

Managers would therefore have a new audience to please. Their jobs would depend less on operational performance, cost control, or consumer demand, and more on keeping the commission satisfied. Research priorities, content-moderation policies, hiring choices, and product decisions would shift toward political approval. The corporation would become less an engine of wealth creation than a government-protected ward with a very expensive engineering department.

The bill’s forced-separation mandate would add another layer of damage. In “Universal Economics,” Armen Alchian and William R. Allen explain that firms combine complementary assets and business lines to reduce transaction costs. A transaction cost is the cost of negotiating, monitoring, enforcing, and coordinating economic activity. Firms often bring activities under one corporate roof because internal coordination can be cheaper and faster than constant contracting among separate companies.

That logic matters acutely in technology. Foundational AI research does not operate in isolation. It depends on hardware procurement, cloud infrastructure, search data, user-interface design, software distribution, and feedback from commercial products. Companies like Alphabet and Microsoft bundle these functions because separating them would require constant negotiation and contracting among units that now coordinate internally.

Forcing AI divisions into legally separate entities would destroy many of those efficiencies. It would replace internal capital allocation and technical coordination with slower, costlier, and more litigation-prone interfirm contracting. That is a particularly bad trade in a sector where speed, iteration, and cross-functional learning often determine whether a product works at all.

The final problem is economic calculation. In “Man, Economy, and State,” Murray Rothbard builds on Mises’ critique of socialism to argue that rational economic calculation depends on private property and genuine market prices for capital goods. Prices are not decorative numbers on a Bloomberg terminal. They summarize dispersed judgments about risk, scarcity, future demand, and alternative uses of capital.

In a pure market, corporate shares and capital inputs are priced through competitive bids by private owners risking their own money. Those prices let entrepreneurs calculate profit and loss. Profit and loss, in turn, guide resources toward uses that consumers value and away from projects that waste capital.

A government-backed sovereign wealth fund with a 50% equity stake in leading technology firms would distort that process. Its investment and voting decisions would not turn on expected consumer demand or profit and loss. They would turn on political mandates, distributional goals, ideological preferences, and pressure from organized groups.

That distortion would affect more than the covered firms. Stock prices help investors value related companies, suppliers, customers, and competing technologies. Once a massive political shareholder begins steering capital and control rights according to non-market criteria, those prices become less reliable guides. The result would be calculational chaos in AI investment, with capital pushed toward politically favored projects and away from technologies that consumers and businesses would otherwise choose.

The $200 Million Tripwire

To fund its redistributive scheme, the AI Sovereign Wealth Fund Act would impose a 50% equity seizure on AI firms that exceed an arbitrary $200 million capitalization or revenue threshold. The bill treats capital as if it were a pile of cash waiting to be divided. Economists sometimes call this the “fundist” fallacy, or the mistaken view that capital is a homogeneous, liquid, and durable fund that can be taxed, sliced, or liquidated without changing the structure of production.

The Austrian theory of capital, developed by Eugen von Böhm-Bawerk and refined by F.A. Hayek, starts with the opposite premise. Capital is heterogeneous. It consists of specific goods, relationships, skills, and investments arranged across time. In artificial intelligence, that means high-performance graphics-processing-unit (GPU) clusters, specialized fiber-optic networks, proprietary algorithms, cloud contracts, data-center capacity, and engineering teams with hard-to-replicate expertise.

Advanced AI development is a long and uncertain production process. Developers often commit billions of dollars years before those investments produce marketable products. Böhm-Bawerk called this “roundabout” production, meaning that firms first invest in earlier-stage capital goods that later help produce consumer-facing goods and services. Those capital goods are not interchangeable Lego bricks. A GPU cluster built for training frontier models, or a team trained to optimize a specific architecture, cannot be painlessly redeployed if the project suddenly becomes uneconomic.

Sanders’ bill would force an artificial restructuring of these capital arrangements. Firms facing the loss of half their corporate value would see their cost of capital rise sharply. Investors would demand higher returns to compensate for confiscation risk, or they would put their money elsewhere. To satisfy the new constraints, developers would scale back, delay, or abandon long-term research projects. Specialized investments in chips, infrastructure, and engineering talent would lose value. The damage would appear not only in projects killed today, but in discoveries never attempted tomorrow.

The bill compounds that problem by shifting control to surrogate decision-makers. In “Knowledge and Decisions,” Thomas Sowell distinguishes between decisions made by people who bear the costs of their choices and decisions made by insulated actors spending other people’s money. In a market, venture capitalists and technology entrepreneurs face a harsh feedback loop. If they misread demand, choose the wrong software architecture, or misallocate capital, they lose money. That risk forces them to revise expectations, correct errors, and shut down failing projects before they consume still more scarce resources.

The seven commissioners managing the AI Sovereign Wealth Fund would face no comparable discipline. They would make centralized bets on the future of computation with other people’s money. If they pushed capital toward a failing model architecture or forced a company to adopt an inefficient safety protocol, they would not bear the financial losses. The usual market signals of profit and loss would weaken. Errors would last longer, spread further, and impose costs on firms, workers, investors, and consumers who had no meaningful say in the decision.

The bill’s $200 million threshold adds another defect. Gordon Tullock’s work on rent seeking shows that firms waste resources when political decisions determine economic rewards. Instead of investing in production, firms invest in protection. They hire lobbyists, seek exemptions, and design their conduct around political risk. The cost includes not only the money spent on lobbying, but also the innovations that never happen because the expected return has been taxed, threatened, or bargained away.

Under Sanders’ bill, an AI startup valued at $199 million remains a private enterprise. Once it reaches $200 million, it faces a 50% equity seizure. That kind of cliff creates perverse incentives. Firms would slow growth, refuse capital, split into smaller entities, or hold back commercial deployment to avoid crossing the line.

The same threshold would redirect scarce talent. Instead of competing to build better models for customers, growing firms would spend more time lobbying the commission for exemptions, waivers, and favorable treatment. Engineers and executives would be pulled away from product development and into regulatory survival. The sector would trade discovery in the lab for supplication in the hearing room. Not exactly the future of artificial intelligence anyone should be rushing to beta test.

The Fatal Conceit, Now With Equity Shares

The American A.I. Sovereign Wealth Fund Act is a near-perfect specimen of what Hayek called the “fatal conceit.” It assumes that a small committee of political appointees can possess the dispersed, tacit, constantly changing knowledge needed to direct a complex technological system. It also reflects what Sowell called “the vision of the anointed,” or the belief that selected experts have both the moral authority and the practical capacity to make high-stakes choices for everyone else.

The bill’s central claim that artificial intelligence is a “public resource” because models train on public data confuses inputs with production. Raw information does not become valuable by existing. Value emerges when entrepreneurs, engineers, investors, and firms discover ways to organize information into tools people actually use. Treating that discovery as public property would weaken the private rights that support innovation in the first place.

The promise of a “democratically directed” AI commission fares no better. Arrow’s Theorem shows why no committee can convert conflicting public preferences into one coherent “social welfare” choice without smuggling in agenda control. Public-choice theory explains what follows. The commission would not embody the “public interest.” It would become a forum for interest-group bargaining, logrolling, and capture by incumbents with the money and lawyers to work the system.

The bill’s structural mandates would deepen the damage. A permanent 50% government voting block would freeze the market for corporate control, weaken managerial discipline, and turn executives toward political survival rather than consumer demand. Forced separation of AI divisions would break apart business structures that firms created to reduce costs, coordinate research, and move quickly. The result would be slower decisions, higher contracting costs, and worse economic calculation..

The bill’s equity seizure also treats capital as if it were a liquid fund waiting for redistribution. High-technology production does not work that way. AI development depends on specific, time-sensitive investments in chips, infrastructure, software, and engineering talent. A confiscatory threshold would raise capital costs, encourage firms to stay small or split themselves apart, and redirect scarce talent toward lobbying rather than building.

Real progress in artificial intelligence does not require central planning by a federal committee. It requires private property, freedom of contract, open competition, and legal rules that let entrepreneurs test ideas, fail cheaply, and scale when consumers find value in what they build.

The future of artificial intelligence will not be discovered by seven commissioners armed with voting shares and a mission statement. It will be discovered by people free to build, buy, reject, improve, and try again.

The state should not claim ownership over artificial intelligence’s future. It should protect the freedom that makes that future possible.

The post The Fatal Conceit Gets a GPU Cluster: Bernie Sanders’ Plan to Socialize AI appeared first on Truth on the Market.

In its June 29 Chatrie v. United States decision, the Court held 6-3 that law enforcement conducts a Fourth Amendment “search” when it forces companies like Google to turn over users’ location-history data through a geofence warrant. 

That is a major win for digital privacy. A geofence warrant lets police demand information about every device in a defined area during a set time—essentially asking first and sorting suspects later. The Court was right to rein in that digital dragnet. 

But peer under the hood of Justice Elena Kagan’s majority opinion, and the legal engine sputters. The Court reached the right destination, but took the scenic route through a swamp. 

To preserve the aging Katz “reasonable expectation of privacy” test—and its creaky cousin, the third-party doctrine—the majority tied itself into a logical pretzel. Justice Neil Gorsuch, concurring only in the judgment, offered the cleaner and more textually grounded path the Court should have taken. 

Keeping Katz on Life Support

To understand the majority’s legal contortions, it helps to start with two pillars of modern Fourth Amendment doctrine: Katz v. United States (1967) and the third-party doctrine. 

Under Katz, government conduct counts as a Fourth Amendment search only if it violates a privacy interest that society recognizes as “reasonable.” The third-party doctrine, which emerged from Katz in the 1970s, adds another wrinkle: if you voluntarily share information with a third party—such as a bank or telephone company—you generally lose any reasonable expectation of privacy in that information. 

In Chatrie, the government leaned heavily on that doctrine. It argued that because Okello Chatrie chose to enable Google’s Location History feature, he voluntarily shared his movements with Google and therefore gave up any Fourth Amendment protection in that data. 

Faced with the prospect of allowing police to obtain the minute-by-minute location histories of hundreds of millions of Google users without a warrant, the Court had two options. It could reconsider the third-party doctrine altogether, or it could conclude that users never truly “voluntarily” shared their data in the first place. The majority chose the latter. 

Writing for the Court, Justice Kagan argued—drawing heavily on Carpenter v. United States—that smartphone location data is different because smartphones have become an indispensable part of modern life. Users do not share their location history in the ordinary sense of wanting Google to have it, the Court reasoned. Instead, disclosing that information is simply “the automatic price of conventional cell-phone usage.” 

As the majority explained: 

[The third-party] argument ignores some pertinent facts about how and why Google users turn on Location History. As described earlier, Google prompts a user, and repeatedly, to turn on the service—when he sets up a Google account, when he sets up an Android phone, and when he sets up a Google app. See supra, at 3–4. The prompt often informs him that his device will not “work correctly” unless he does so. 2 App. 140–141. By contrast, it does not tell him quite what he is signing up for: “how frequently Google would record [his] location”; “how precise Location History can be”; or how Google might give all that minute-by-minute location information to the government. 590 F. Supp. 3d, at 936; 136 F. 4th, at 128 (Wynn, J., concurring in judgment). In those circumstances, it is hard to see how any user is, in the normal sense, “sharing” with third parties a comprehensive catalog of his physical movements. Carpenter, 585 U. S., at 314. 

That reasoning has implications well beyond geofence warrants. It effectively casts doubt on whether consumers meaningfully consent to the standard-form contracts that govern modern digital services. In today’s economy, users routinely exchange data for free or discounted access to hardware, software, and online services. If that exchange is not truly voluntary, the legal foundation for countless digital transactions becomes far less certain. 

The Court’s rationale depends on precisely that premise. It notes that more than 90% of Americans own smartphones and suggests many are effectively dependent on apps and services that collect and store “detailed information about all aspects of a person’s life.”

The result is a curious theory of privacy. Under the majority’s approach, you retain a reasonable expectation of privacy in data that you knowingly allowed a global technology company to collect, record every few minutes, and store on its servers because modern life and take-it-or-leave-it contracts left you with little practical choice. By recasting ordinary participation in the digital economy as a form of structural coercion, the majority keeps Katz alive only by carving out a highly subjective, ad hoc exception to the third-party doctrine.

The Property Path Not Taken

Justice Gorsuch’s concurrence points to a better path. He agrees that the government conducted a Fourth Amendment search requiring a warrant, but rejects the Katz framework altogether. Instead, he asks a much simpler question grounded in the Constitution’s text: Is your digital location data your property? Is it one of your “papers” or “effects”? 

Gorsuch argues that the Katz test has “no basis in the Constitution’s text or history,” relying instead on the shifting and unpredictable intuitions of judges. He similarly criticizes the third-party doctrine as an artificial rule that departs from traditional property law. 

In everyday life, entrusting your property to someone else does not mean surrendering your rights to it. As Gorsuch explains, when you “toss your keys to a valet at a restaurant” or “ask your neighbor to look after your dog while you travel,” you have entrusted your property to someone else. You have not abandoned it, exposed it to the public, or given the police permission to rummage through your trunk or seize your pet. 

The law has long recognized this arrangement as a bailment—a legal relationship in which one person temporarily entrusts property to another while retaining ownership and the right to exclude everyone else. Gorsuch argues that storing data with Google is no different. 

Why should digital property receive less protection than physical property? Chatrie’s Location History is, in effect, an electronic diary or a detailed map of his movements. He retains the right to use it, modify it, delete it, and exclude others from it. That includes contracting with Google to store the data under an agreement promising to protect it from “unauthorized access, alteration, disclosure, or destruction.” 

By grounding the Fourth Amendment in property rights rather than amorphous expectations of privacy, Gorsuch avoids the majority’s logical trap. If Chatrie’s location history is one of his digital “effects”—his personal property—the government cannot search or seize it without a valid warrant. It makes no constitutional difference that Google stores the data, just as it would make no difference if Chatrie kept his private journals in a rented storage unit. 

The Right Result, the Wrong Rule

Some might argue that once the Court required a warrant for geofence warrants, the underlying legal theory became an academic debate. It did not. The majority’s convoluted reasoning leaves our digital rights on shaky ground. 

Under Katz, those rights depend on whether five Supreme Court justices decide a particular app, device, or service is sufficiently “compulsory” or “revealing” to deserve constitutional protection. That invites years of feature-by-feature litigation. The government will continue arguing that users “voluntarily” disclosed their information by adopting the next new technology, or that whatever technology is at issue is not truly indispensable to modern life. 

A property-based approach offers a sturdier foundation. It does not ask judges to psychoanalyze why people use smartphones or speculate about whether consumers had a meaningful choice. Instead, it asks whether an individual retained a legal right to exclude others from the information. 

That principle is hardly novel. The law has long recognized that people can protect their property through practical measures, such as locking a door or building a fence, and through legal arrangements, such as contracts or bailments. Properly understood, many of the values underlying the “reasonable expectation of privacy” test can be preserved by grounding them in longstanding principles of property and contract law rather than in evolving judicial intuitions. 

Chatrie v. United States is a landmark victory for digital liberty. But the next generation of constitutional disputes—many involving artificial intelligence and technologies that do not yet exist—will demand more than the right result in one case. They will require the right constitutional framework.

If the Fourth Amendment is to protect our digital “papers” and “effects” as faithfully as it protects their physical counterparts, the Court should stop stretching Katz to fit the digital age and start taking the Constitution’s text at its word.

The post Chatrie and the Court’s Pretzel Logic: The Fourth Amendment Gets Twisted appeared first on Truth on the Market.

The Commission has reached the preliminary view that Amazon Web Services (AWS) and Microsoft Azure should be designated as gatekeepers under the DMA. That would mark two firsts. It would be the first time cloud computing falls within the DMA’s reach, and only the second time the Commission uses Article 3(8)—the provision that allows it to designate firms that do not meet the law’s numerical thresholds after conducting a market investigation. While the Commission did invoke Article 3(8) in its iPad OS designation, quantitative thresholds still did most of the work in that case.

AWS and Azure are the two largest cloud providers operating in the European Union. But the DMA’s user-number thresholds were built for consumer-facing platforms, not cloud computing, which is overwhelmingly a business-to-business service. Both providers therefore fall outside those thresholds. To designate them anyway, the Commission must do more than invoke the DMA’s built-in presumptions. It must show, with actual evidence, that AWS and Azure serve as “important gateways” and enjoy “entrenched and durable” market positions.

That makes these designations far more interesting than another lap around the DMA enforcement track. Cloud services do not obviously operate as “gates” in the way two-sided platforms do. They do not sit between business users and end users in the familiar app-store or marketplace sense. One of the DMA’s core rationales—increasing contestability by prying open bottlenecks controlled by gatekeepers—appears, at least at first glance, to be missing.

So the question is not merely whether AWS and Azure are large. Plainly, they are. The question is whether “gatekeeper” remains a meaningful legal category that separates firms with durable bottleneck power from firms that are big in competitive markets. Or is it just Brussels-speak for size, with the statutory criteria serving as the ceremonial chant before the inevitable designation?

Until now, the Commission has relied on the quantitative presumptions in Article 3(2) to designate gatekeepers. The qualitative criteria have never had to carry the load on their own. With AWS and Azure, they finally do.

Can Cloud Be a Gate if No One Walks Through It?

It is tempting to argue that cloud computing simply does not belong in the DMA at all. After all, cloud is infrastructure. It rarely touches consumers directly. On that view, it cannot be an “important gateway for business users to reach end users.” 

As a matter of market reality, that intuition has some force. As a matter of law, it is much harder to sustain. 

Cloud computing appears expressly in Article 2 as a core platform service. Recital 14 explains why: the listed services have “the capacity to affect a large number of end users and business users, which entails the risk of unfair business practices,” and therefore should fall within the DMA’s scope. 

That verb matters. Recital 14 does not say cloud providers “have” a large number of end users. It says they have “the capacity to affect” them. Had the drafters meant the former, they could have said so. The qualitative gateway criterion in Article 3(1)(b) would then do no independent work; it would collapse into the numerical test in Article 3(2). 

The choice of “affect” suggests the legislature contemplated—and accepted—that a service could be designated even when it reaches end users only indirectly, through the businesses that build on top of it. In other words, the qualitative thresholds exist precisely for services like cloud, which can affect end users without serving as their direct intermediary. 

But “can be designated” does not mean “should be designated.” Eligibility says nothing about whether these providers, in this market, actually satisfy the substantive criteria. Indeed, the very reason the DMA’s recitals make cloud eligible—that it can affect end users without directly intermediating with them—should put us on guard. 

When a core platform service lacks the usual features that support a finding of gatekeeper power—when it is not a traditional multi-sided, consumer-facing platform—the remaining criteria deserve more scrutiny, not less. The key question is whether those criteria establish a genuine competitive bottleneck, or merely dress up firm size in statutory clothing. 

Size Is Only the First Box

Start with the structure of the test. Article 3(1) makes designation depend on three cumulative conditions. The undertaking must have a “significant impact on the internal market.” It must provide a core platform service that serves as an “important gateway” for business users to reach end users. And it must enjoy an “entrenched and durable position,” either now or in the foreseeable near future. 

“Cumulative” is doing real work here. Each element must stand on its own. Evidence that supports one cannot simply be recycled to prop up another. A decision that lists indicators of scale—turnover, capitalization, investment, and vertical integration—and then presses them into service under every heading has really established only the first condition: that the company has a “significant impact on the internal market.” 

For AWS and Azure, that first condition is not seriously in dispute. Firms of their size clear it easily. The fight is over the two remaining criteria: whether they are “important gateways,” and whether their positions are “entrenched and durable.” 

Big Is Not a Bottleneck

The DMA’s gateway concept is built around intermediation. A core platform service sits between business users and the end users they want to reach. As the General Court held in ByteDance: 

[I]n order to consider that business users of a CPS [core platform service] “depend” on it in order to reach their end users, it is not necessary for that CPS to be the only channel through which those undertakings can reach those users. It is sufficient for it to be an important channel for that purpose, which those business users can access only if they have an account on that CPS. (ByteDance, §210)

That distinction matters. The first and third limbs of Article 3(1), taken alone, describe nothing distinctively digital—and nothing distinctly gatekeeper-like. Plenty of firms have a “significant impact” on the internal market. Plenty occupy “entrenched and durable” positions. What separates a gatekeeper from a company that is merely large and long-established is the middle limb: whether the firm provides an “important gateway” for business users to reach end users.

Take that limb seriously, and the DMA has a subject matter of its own: bottlenecks that let certain firms control how business users meet their customers. Hollow it out, and the regulation becomes a checklist of obligations for companies that have been big for three years running.

This gateway model fits marketplaces, social networks, and app stores. It fits raw compute and storage much less comfortably. A business that hosts its application on AWS does not reach customers “through” AWS in any way those customers see or experience. Cloud computing is, instead, an input into the business’s own offering.

Article 3(1)(b) contains two distinct requirements. The service must be a gateway, and the gateway must be important. The presumption that ordinarily carries this limb—Article 3(2)(b)’s thresholds of 45 million monthly active end users and 10,000 yearly active business users—does double duty. Both numbers matter because the legal concern is intermediation between two groups.

The size of the numbers speaks to importance. A service reaching users at that scale plainly matters to the internal market. But the structure of the threshold—the requirement of large numbers on both sides, business users and end users—speaks to the gateway element. A service connecting 10,000 businesses to 45 million consumers is, almost by definition, intermediating between the two groups.

For cloud, both halves of that proxy are missing, and for the same reason. The end users are indirect. People who use an app usually have no relationship with the infrastructure on which it runs. So the numbers are not there. And because cloud has no second side in the relevant sense, the Commission cannot infer “gateway” from “many users of both types.”

That is what forced the Commission into Article 3(8). But Article 3(8) comes with a price: the Commission must prove both “important” and “gateway” separately, in substance, without leaning on the presumption. Evidence that cloud is big—revenues, capacity, investment, or ubiquity as an input—may establish importance. It says much less about gateway status. A decision that stacks up evidence of size and criticality and calls the pile an “important gateway” has proved only half the statutory phrase. It has skipped the half that makes a gatekeeper a gatekeeper, rather than merely big, significant, or useful. 

This should not be controversial. The Commission declined to designate X despite X clearing the DMA’s turnover thresholds precisely because it was not an important gateway. It accepted that Gmail and Outlook were not gatekeepers despite each having many millions of business and end users. iMessage cleared the quantitative thresholds and was nonetheless found not important enough. 

Those decisions concede that the words in Article 3(1)(b) matter. Designating cloud—where the gateway-to-end-users fit is at its weakest—therefore creates a dilemma. Either the Commission is applying a more lenient gateway standard to infrastructure than it applied to X and email services, raising an equal-treatment problem the firms can and likely will plead; or it is conceding that, for infrastructure, “gateway” now means “big.” 

There is, admittedly, a reading of the gateway limb that keeps it meaningful while leaving cloud eligible in principle. It is also the reading to which the Commission should be held. On this view, “gateway” captures the market structure that motivated the DMA in the first place: platform intermediation between distinct user groups, where network effects, data feedback loops, and extreme returns to scale can tip markets and lock them in. That was the diagnosis of the Crémer report and the Furman review, and it is the theory on which the regulation was built. 

If that is what “gateway” means, a cloud designation is not impossible. But it requires the Commission to show that something about this market inherently tends toward tipping and entrenchment. That leads directly to the second contested limb, where the evidence can actually be inspected. 

Entrenched? Tell That to the Price Cuts.

Here, the economic reality is observable—and it points the wrong way for the Commission. 

Start with the structure of the DMA. Article 3(2)(c) presumes an “entrenched and durable position” when the user thresholds in Article 3(2)(b) have been met in each of the last three financial years. In the DMA’s own scheme, entrenchment means gateway numbers sustained over time. When those numbers are unavailable—and their absence is what pushed the Commission into Article 3(8) in the first place—something else must fill the gap. The only plausible substitute is evidence about contestability itself: whether rivals can challenge the firms’ positions. 

The case law points the same way. The DMA sets aside the dominance test under Article 102 of the Treaty on the Functioning of the European Union (TFEU), which governs abuse of dominance. But the two inquiries are not unrelated. Dominance is a legal test aimed at an economic question: whether rivals can discipline a firm if it raises prices, restricts output, or lets quality and innovation slide. The General Court defines an “entrenched and durable” position in similar terms, referring to situations where “the contestability of that position is limited” and looking to “the stability of that position over time” (ByteDance, §§296-297). 

That makes entrenchment an economic question. The Commission must show actual obstacles to contestability. Strip out that economic reality, and “entrenched and durable” becomes another word for “important,” which the DMA treats as a separate inquiry, or simply for “big.” To designate AWS or Azure on this limb, the Commission must show that their positions are, to some meaningful degree, insulated from competition. 

That raises a broader point. As Chad Syverson put it, concentration “is an outcome, not an immutable core determinant of how competitive an industry or market is”—so much so that “we cannot even generally know which way the barometer is oriented.” 

The same logic applies to the DMA’s qualitative test. Outside the DMA’s presumptions, “entrenched and durable” is not a structural fact one can read off a market’s shape. It is the result of limited contestability. The question is not how concentrated the market looks in a snapshot, but whether competitive pressure can still discipline the incumbents. 

That is why entrenchment cannot rest on market shares alone. It also cannot credibly describe a market where rivals are actively winning and losing business. Yet that is exactly what the cloud industry shows. 

AWS’ market share has, by all accounts, declined while Microsoft and Google have made large gains. Meta is reportedly exploring ways to turn unused capacity from its massive Meta Compute buildout into a cloud business, potentially by hosting AI models or selling bare-metal capacity—dedicated server capacity without a cloud provider’s software layer—to so-called neocloud providers, which are newer cloud firms focused largely on AI workloads. 

Prices have also fallen dramatically. Amazon Simple Storage Service (S3), AWS’ basic data-storage service, dropped more than 80% in price in the early 2010s and has largely stagnated since, despite rising inflation. Tellingly, AWS accelerated its price cuts once Microsoft entered at sufficient scale to post competitive prices. That is not the behavior of a firm floating above the market. It is rivalry doing its work. 

The very practices the Commission treats as lock-in mechanisms have also loosened under competitive and regulatory pressure. AWS, Azure, and Google have scrapped or slashed egress fees—charges customers pay to move data out of a cloud provider’s system—for customers leaving their platforms. A firm that keeps lowering the switching costs that supposedly trap its customers is demonstrating, in real time, that its position is contestable. Around 70% of cloud users also multi-home, meaning they use more than one cloud provider rather than putting all their eggs in one server rack. 

Article 3(8) does not let the Commission paper over any of this. Removing the quantitative thresholds means the Commission must prove limited contestability in substance. And the substance here is a market where prices fall, shares move, customers multi-home, and incumbents cut their own switching costs to keep business. 

There is another problem the Commission’s reasoning must survive: aggregation. The figure most often cited for cloud—a combined share north of 50%—is the share of two firms together. But the DMA’s test applies at the undertaking level, and Article 3(9) requires each designated service to be a gateway “individually.” 

There is no equivalent of Article 102 TFEU’s “collective dominance” under the DMA. A combined number tells us little about whether AWS or Azure individually occupies an unassailable position. Instead, it lumps together two companies that compete intensely with each other and with other providers.

The Commission’s decision to designate AWS and Azure while leaving Google Cloud outside the gate makes the problem sharper. Without the thresholds, what principled line puts Amazon and Microsoft in, but Google out? If the honest answer is “Google has fewer users,” then the unstated basis for the whole exercise is user numbers after all—which is to say, bigness. Microsoft and Amazon are right to press this point: there is no obvious, legitimate basis for excluding Google while including them that does not collapse into the very head-counting Article 3(8) was supposed to move beyond.  

This is not to say Google should be designated. It is to say that the apparent basis for designating AWS and Azure is that they have acquired an arbitrarily “high” number of users. That number is not high enough to trigger the presumptions in Article 3(2), but it is apparently high enough to support an inference of “significant impact.” If the whole designation turns on user numbers anyway, what exactly are the quantitative presumptions for? 

One caveat remains. The Commission may answer that the problem is forward-looking. The DMA allows it to designate a firm when it is “foreseeable that it will enjoy” an entrenched position “in the near future.” The Commission could concede that AWS and Azure face real competition today, but predict that AI workloads, “data gravity”—the tendency of applications and services to cluster around large pools of data—and sovereign-cloud demand will cement their positions tomorrow. 

But “foreseeable entrenchment in the near future” is still a claim about contestability. It requires concrete evidence that competition will stop constraining the firms. The General Court’s Meta ruling enforces that standard. 

In that case, the Court annulled the Marketplace designation, not after a full substantive reassessment, but because the Commission’s reasoning about one factual change was “vague and hypothetical,” offered no “specific analysis,” and because the legality of a designation must be assessed “on the basis of the facts and the law as they stood at the time when the measure was adopted.” 

A near-future-entrenchment finding built on “cloud is large and AI will lock things in” is conclusory in exactly that way. The forward-looking limb does not license regulatory fan fiction. It requires evidence. 

There is also some irony in invoking AI to predict cloud entrenchment. Cloud is precisely what has democratized AI for smaller firms, putting frontier models within reach of startups through services such as Amazon Bedrock and Azure’s OpenAI offering. The dynamic the Commission points to as tomorrow’s bottleneck is, so far, helping lower the barriers today. 

The DMA Without Its Training Wheels

None of this guarantees that potential cloud designations will fall. Article 3’s relationship between quantitative presumptions and qualitative criteria allows two readings. On the restrictive reading, the qualitative criteria serve as a real filter. On the permissive reading, they are broad enough that the Commission can “always find something” once a firm reaches sufficient scale. 

Because the DMA is written at a high level of generality, the honest prediction is that the Commission will choose the permissive route—and the courts may let it. The regulation does not clearly foreclose that reading. Brussels left itself a wide lane, and it should surprise no one if it drives down the middle of it. 

The European Court of Justice’s early rulings suggest any judicial check will not come from a court-imposed market-power standard. Recital 5 cuts against that. Instead, it will come—if it comes at all—from careful review of the Commission’s reasoning.

ByteDance and Meta were both presumption cases. The Commission won on substance because the thresholds, not its own independent analysis, carried the qualitative criteria. The one time the General Court struck down a designation—Marketplace—it did so because the Commission’s reasoning failed. The Court held the Commission to the facts as they stood when the designation was adopted and demanded a concrete analysis the Commission had not supplied.

The standard is the ordinary one for any EU act:

[T]he statement of reasons required by … Article 296 TFEU … must disclose in a clear and unequivocal fashion the reasoning followed by the institution which adopted the measure … to enable the persons concerned to ascertain the reasons for it and to enable the court having jurisdiction to exercise its power of review. (Meta, §53)

In other words, the European Court of Justice is unlikely to say “market power.” But it can vindicate the same substance by requiring the Commission to explain, with evidence rather than adjectives, why a sub-threshold service is both a gateway and entrenched. Whether the qualitative criteria mean anything will depend on whether the courts insist that the Commission do more than admire the size of the target. 

That is why the AWS and Azure designations matter so much. In every prior designation, the presumption let the hardest questions go largely untested. The courts could avoid asking what “important gateway” and “entrenched and durable” mean when the numbers do not carry the answer. They cannot duck those questions here. 

If the Commission designates AWS and Azure, appeals are almost certain. The resulting litigation may be the most consequential the DMA has yet produced. Article 3(8) offers no numerical crutch. The Commission must prove gateway power and entrenchment on a full record, under the kind of reasoning review the General Court applied to Marketplace. 

As the General Court has confirmed, Article 3(8) is the route for examining “whether an undertaking should be designated pursuant to Article 3(8)… despite the fact that the undertaking or the [core platform services] in question do not satisfy the thresholds.” Put differently, these two designations would be the first time the DMA’s substantive criteria must stand on their own. 

They will also answer the question this post began with. If a sub-threshold cloud designation survives on “large and sticky”—in a market with falling prices, shifting shares, widespread multi-homing, and incumbents cutting their own switching costs under competitive pressure—then the answer is settled. The DMA is a test of bigness. There was nothing else available to carry the designation, and bigness carried it anyway. 

The contrast with the parallel exercise across the Channel makes the stakes even clearer. The United Kingdom’s Competition and Markets Authority (CMA) examined the same market and the same two firms under a regime whose gatekeeper logic closely tracks the DMA’s. Its market investigation found that AWS and Microsoft hold significant market power, and the inquiry group recommended that the CMA Board prioritize strategic market status (SMS) investigations into both providers’ cloud businesses. 

The Board looked at the same record and—so far, at least—declined. In March 2026, it opted instead for supervised commitments. The one concern the CMA deemed serious enough for possible designation—Microsoft’s software licensing—was routed into an SMS investigation of Microsoft’s business-software ecosystem, not cloud itself. 

So the divergence is not about the facts. It is about what the facts warrant. Armed with designation powers of its own and a dedicated market-investigation record, the UK authority concluded that supervised commitments—not designation—were the proportionate response to a market that is still moving, and whose centrality to AI-driven growth counsels against regulatory overkill. 

The Commission now proposes to go further on a thinner basis: a first-ever sub-threshold designation, carrying the full weight of the DMA’s obligations, in the very market its British counterpart just concluded could be steered with a lighter touch.

If that designation is made, and if it survives review, the divergence will tell us less about cloud than about the DMA. One regime asked whether these firms control a gate. The other may decide it is enough that they are very large and standing nearby. In Brussels, that may be all the gate you need.

The post The DMA’s Cloud-Cuckoo Land appeared first on Truth on the Market.

California, naturally, has other ideas.

Much of the state’s copper network remains in place—powered, maintained, and protected by rules written for a monopoly telephone era that no longer exists. AT&T still provides old-fashioned “plain old telephone service” (POTS) to roughly 3% of households in its California territory. Yet the company spends about $1 billion a year keeping that network alive.

The reason is not consumer demand. It is state regulation. And the costs do not stop at California’s border. They fall on customers across the country.

The Federal Communications Commission (FCC) now has a chance to fix that. AT&T has asked the FCC to declare that, once the agency authorizes POTS discontinuance under federal law, California’s carrier-of-last-resort (COLR) requirements are preempted. 

The Hotel California of Telecom

California’s COLR regime dates to the mid-1990s, when the state opened local phone markets to competition but worried some customers might get stranded in the transition. It designated incumbent carriers as “carriers of last resort,” obligating them to provide “basic service” to every residential household, and to serve every business on request, throughout their territories.

The state calls these rules “technology neutral.” In practice, they force incumbents like AT&T to keep legacy copper networks running. The required elements of “basic service”—directory assistance, white-pages listings, free operator services, and voice calls “over all distances”—come straight from the POTS era. Modern wireless, cable, and voice-over-internet-protocol (VoIP) providers need not bundle those features into their offerings.

Worse, the California Public Utilities Commission (CPUC) has refused to let a carrier satisfy its COLR obligation with any substitute technology unless the agency first adopts service-quality standards for that substitute. For mobile wireless, it never has. The practical effect is that AT&T can meet its COLR duty only by continuing to operate the same copper network the FCC would let it retire.

Nor is there an easy exit. CPUC rules allow a carrier to relinquish COLR status only if another carrier volunteers to take it on. When more than 200 carriers were offered the chance to assume AT&T’s duties, every one declined. Shocking, yes: No firm wanted an unfunded mandate to run obsolete infrastructure.

When AT&T applied in 2023 to relinquish the designation, the CPUC spent more than a year on contested proceedings, discovery, and statewide hearings. It then dismissed the application on a threshold motion and barred AT&T from reapplying for a year.

California also layers on procedural burdens that independently block incumbent carriers from retiring POTS. State law requires basic service to be tariffed—meaning offered under rates and terms filed with the regulator—and bars carriers from removing it from those tariffs. Any change requires a formal CPUC proceeding the agency can reject.

Discontinuing POTS also triggers customer-notice approvals and a 19-point “exit plan” under the state’s Mass Migration Guidelines, with service required to continue until the agency approves each submission. Each step needs separate sign-off. None carries a deadline the CPUC must honor. Any one can be denied.

The result is a practical veto over modernization—even after the FCC says yes.

The High Cost of Hanging On

It is tempting to dismiss this as California’s problem. It is not.

AT&T and other communications providers do not budget on a state-by-state basis. They allocate capital across nationwide networks, and capital is finite. Every dollar AT&T must spend maintaining aging copper loops and switching offices in California is a dollar it cannot invest in fiber or wireless infrastructure elsewhere.

The numbers are striking. AT&T spends roughly $6 billion a year—about 5% of its revenue—operating a legacy network for a customer base that shrinks every month. California’s COLR mandate locks a significant share of that spending into obsolete infrastructure. Consumers nationwide bear the opportunity cost when faster, more reliable networks arrive later, or never arrive at all.

The market case for retiring copper is overwhelming.

Consumers have already moved on. Copper last-mile subscribers fell about 81% between 2014 and 2024, from roughly 66 million to 12.5 million. By 2024, 79% of U.S. adults lived in wireless-only households, while fewer than 1% relied exclusively on landlines.

Fiber is also dramatically cheaper to operate. Verizon’s migration of 4.5 million circuits to fiber saved roughly $180 million annually and reduced maintenance dispatches by about 60%. All-fiber networks cost about $91 less per home each year than copper-based digital subscriber line, or DSL, networks. Much of those savings come from lower energy use. Altafiber, for example, reported that copper service uses about 172 kilowatt-hours per subscriber annually, compared with just 6 kilowatt-hours for fiber—a 97% reduction.

The equipment is also nearing the end of its useful life. The Lucent 5ESS, Nortel DMS-100, and Siemens EWSD switches that anchor many copper networks have not been manufactured in decades. Carriers increasingly scour secondary markets for replacement parts. In one memorable example, Tinker Air Force Base kept a 5ESS switch running by buying components on eBay.

Copper has also become a lucrative target for thieves. As copper prices climbed from about $2.29 per pound in 2020 to nearly $6 by early 2026, theft surged. AT&T reported about 8,700 theft incidents in 2025, costing roughly $76 million. Across the industry, more than 15,500 theft-and-sabotage incidents in a single year disrupted service for more than 9.5 million customers, including 911 systems, hospitals, and military bases. Requiring carriers to keep valuable copper in the ground long after customers have abandoned it effectively subsidizes organized theft.

The upside of completing the transition is equally substantial. One Brattle Group estimate projects that finishing a nationwide fiber buildout would generate roughly $3.24 trillion in net present value and support about 380,000 jobs. That estimate reflects a complete national transition, not California’s rules alone. Still, it shows the scale of the investment opportunities foreclosed when billions of dollars remain tied up maintaining obsolete infrastructure.

Most importantly, retiring copper does not leave consumers without telephone service. Voice communications have undergone the kind of creative destruction that characterizes competitive markets: newer, better technologies have displaced the switched-access copper network, and consumers have overwhelmingly embraced them.

The FCC’s own data tell the story. Hundreds of millions of mobile subscriptions and tens of millions of interconnected VoIP lines now dwarf a residential switched-access network that continues to shrink at double-digit annual rates.

These alternatives do more than replicate POTS. They offer faster speeds, lower latency, greater reliability, and the ability to combine voice, video, and messaging in a single internet-based service. Recognizing that reality, the FCC’s 2026 Network Modernization Order permits carriers to discontinue copper service only when an adequate replacement is available. In other words, the consumer-protection function California’s COLR rules supposedly serve is already provided under federal law by the same order that authorizes carriers to retire their copper networks.

Where the Wires Cross

That leaves a legal collision.

The FCC’s 2026 Network Modernization Order streamlined the process for retiring legacy networks and expressly preempted state laws that prevent carriers from discontinuing interstate and jurisdictionally mixed services. “Jurisdictionally mixed” simply means a service that carries both interstate and intrastate traffic, as modern communications networks often do.

The order also rests on a straightforward statutory command. Section 214 of the Communications Act provides that, once the FCC authorizes a discontinuance, a carrier may proceed “without securing approval other than such certificate.”

California demands precisely the additional approvals that Section 214 appears to forbid. The complication is that the FCC’s authority under Section 214 does not extend to services that are purely intrastate—communications that begin and end within one state.

In theory, that distinction is clean. In practice, modern networks have made a mess of it. Voice calls, data packets, mobile connections, and VoIP services do not always respect the tidy jurisdictional boxes designed for a simpler telephone system. Whether the FCC can preempt California’s COLR regime therefore turns on the nature of the service at issue.

Even when a state insists it is regulating only intrastate service, federal law recognizes an important limit. Under the long-established “impossibility exception,” courts have held that federal law may preempt state regulation when interstate and intrastate components cannot realistically be separated—not only because they are physically intertwined, but because separating them would be practically or economically impossible.

That is the question the FCC now faces. If California’s COLR rules operate as a backdoor veto over federally approved discontinuance of interstate or mixed services, federal law should control. States may protect consumers. They may not use yesterday’s consumer-protection rules to override a federal decision that yesterday’s network can finally be retired.

The Last Call for Copper

California’s COLR rules were designed to protect consumers during the transition from monopoly telephone service to competitive markets. That transition happened years ago. What remains is a permanent tax on modernization—one that singles out a single provider, protects few consumers whom the market has not already reached, and quietly reduces the capital available to build fiber and wireless networks across the country.

The FCC likely has the legal authority to end this stalemate. The economic case for doing so is even stronger.

When the federal government has determined that a century-old network can finally be retired, no single state should be able to force it to live forever. At some point, even the last copper wire deserves to be disconnected.

The post The State That Wouldn’t Hang Up: California’s Fight to Keep the Old Phone Network Alive appeared first on Truth on the Market.

In the first article in this series, we examined how Brazil’s Administrative Council for Economic Defense (CADE) has approached this new generation of artificial-intelligence partnerships. We focused on the agency’s treatment of “reverse acqui-hires”—deals in which a company hires much of a startup’s team without formally buying the company—as concentration acts; its renewed reliance on Article 88, §7 call-in powers, which allow CADE to review certain transactions that fall outside Brazil’s mandatory-notification thresholds; and the practical distinctions that pushed Microsoft/Inflection—but not Google/Character.AI—into formal merger review. 

Those procedural moves raise a deeper question: What theory of harm is driving CADE’s enforcement agenda?

CADE’s decisions suggest that the agency is increasingly borrowing from international debates over “killer acquisitions,” “talent hoarding,” and the loss of potential competition in digital markets. A “killer acquisition” occurs when an incumbent buys a smaller firm to shut down a potential future rival. “Talent hoarding” is the related claim that large firms hire scarce technical workers not to deploy them productively, but to keep them away from rivals. These theories sound tidy enough. The harder question is whether they rest on evidence sturdy enough to support intervention.

Before CADE completes its review of Microsoft/Inflection or opens further investigations into AI partnerships, that question deserves a closer look. This second article takes it up.

The Theory Comes Before the Evidence 

At CADE’s May 13 court session, interim President Diogo Thomson gave an oral “preamble” statement that the agency later incorporated into the decisions. He warned that “transactions in the digital, technology, and AI markets may lead to significant exchanges of assets, capabilities, technology, intellectual property, key employees, and competitive advantage, even if the companies involved have no significant turnover in Brazil.” In those cases, he said, CADE “performs as an outstanding mechanism to fix the mandatory notification system, exhausting all possibilities of potential competition concerns within the legal limits.” 

That is the language of “killer acquisitions” and “acqui-hires,” with a visible assist from the 2024 Joint Statement by the U.S. Department of Justice (DOJ), Federal Trade Commission (FTC), European Commission, and the United Kingdom’s Competition and Markets Authority (CMA) on competition in generative-AI foundation models and AI products, as well as from the CMA’s Microsoft/Inflection inquiry decision. Both Commissioner Camila Alves and Commissioner José Levi Mello do Amaral Júnior anchor their reasoning in this frame. 

Alves identifies four possible theories of harm related to AI partnerships: (i) the loss of potential competition and abandonment of an independent innovative path, which “engage in dialogue with the killer-acquisition literature, although they are not entirely synonymous with it” (§80); (ii) the absorption of organized technical capacity; (iii) the reinforcement of ecosystem integration and asymmetric exploitation of technology licenses; and (iv) potential effects on specialized talent markets and innovation incentives—the so-called “talent hoarding” theory.

Amaral’s opinion, which led to the order requiring formal notification of Microsoft/Inflection, goes further. He writes that “there is a growing understanding that reverse acqui-hire transactions may constitute killer acquisitions” (§67) and that the record contains sufficient evidence that the transaction is “highly likely to give rise to anticompetitive effects” (§69). The killer-acquisition theory thus became the main doctrinal basis for the call-in decision. It deserves scrutiny.

Start with the evidence. The original paper by Colleen Cunningham, Florian Ederer, and Song Ma found that roughly 5%-7% of acquisitions resulted in discontinuation of the target’s development projects. But that evidence came from the pharmaceutical sector, where drug pipelines, patent rights, regulatory approvals, and product-development timelines look very different from software and AI markets. As Selçukhan Ünekba? has cautioned, those findings cannot simply be airlifted into digital markets and treated as proof of the same problem. 

In an Antitrust Law Journal article, Marc Ivaldi, Nicolas Petit, and Ünekba? find no support for the killer-acquisition theory in European digital-merger cases: 

We focus on transactions reviewed by the European Commission in information and communication technology industries. Pursuant to the theory of killer acquisitions, some of these cases should have led to reduced competition. Focusing on publicly available information through financial disclosures, our analysis suggests that no transaction was followed by the disappearance of the target’s products, a weakening of competing firms, and/or a post-merger lowering or absence of entry and innovation. Skepticism about the killer acquisitions theory should prevail.

Even Ederer, one of the original killer-acquisition authors, has complicated the story. In a recent working paper with Reginald Seibel and Timothy Simcoe titled “Digital (Killer?) Acquisitions,” the authors examine “the effects of 1,200 acquisitions by major technology firms on innovation.” Their conclusion is worth quoting at length: 

[O]ur findings complicate the popular narrative that digital acquisitions by large incumbents are predominantly anti-competitive. While we do not rule out the possibility of killer acquisitions in individual cases, our evidence suggests that most deals by GAFAM and related firms are motivated by complementarity rather than suppression. Innovation tends to rise rather than fall in the wake of acquisition, particularly in technological domains where the acquirer has prior experience and continues to make follow-on investments.

Here, GAFAM refers to Google, Apple, Facebook (now Meta), Amazon, and Microsoft. The point is not that acquisitions by large technology companies can never harm competition. The point is narrower, and more important: the evidence does not support treating such deals as presumptively suspect. 

The “talent hoarding” variant rests on similarly fragile foundations. Geoffrey Manne recently dissected a paper by Shaolong Wu of Harvard Business School and Zefan Qian of Georgetown, “Talent Hoarding and Upstream Innovation: Labor Market Distortions by Large Incumbents,” that has been offered in support of the theory. Manne shows that its headline claims of “idle benches,” foreclosed rivals, and lost innovation follow from the model’s assumptions, not from the data. The data are also consistent with a more prosaic explanation: firms efficiently retain employees who embody proprietary know-how. 

Manne also notes that today’s AI incumbents are doing the opposite of what the hoarding story predicts. Between 2019 and 2025, the largest technology firms added nearly 1 million employees while increasing capital expenditures from $77 billion to $370 billion. That looks less like firms benching talent and more like firms racing to deploy it.

Importing these theories into Brazilian case law on the back of a still-forming literature—one that, for now, often points the other way—is hardly a prudent course.

Acqui-Hires Aren’t Mini-Mergers 

A merger usually transfers durable business assets: companies, product lines, customer relationships, intellectual property, and control. An acqui-hire mainly transfers people. That distinction matters because people are not factories. They can leave, start companies, join rivals, publish papers, and take their know-how with them. 

Onyeka Aralu and Dirk Auer explain the point well. Talent transfers are temporary because workers can leave. Team acquisitions are often more transient and speculative than the durable structural changes merger law targets. AI-talent supply is expanding, not fixed. And target companies that retain their intellectual property and customers can, and often do, keep operating, sometimes with a new strategy. 

Ünekba? has also cataloged the legitimate efficiency reasons for these deals. They may help a firm reposition strategically, avoid the high information costs of assembling a cohesive team one hire at a time, or manage the failure of promising but struggling startups. In plainer terms: Sometimes a team deal is not a plot. Sometimes it is a way to salvage talent, technology, or both. 

AI’s “moats” are also unusually permeable. In business jargon, a moat is a durable competitive advantage that protects a firm from rivals. But AI often works differently. As one of us has explained elsewhere, the field’s scientific ethos—open publication, mobile talent, and paper-driven diffusion—means the “capabilities” transferred in an acqui-hire often do not remain proprietary for long. 

Google’s 2017 Transformer paper is the classic example. By publishing “Attention Is All You Need,” Google released the architectural blueprint for much of the generative-AI boom, while several of the paper’s authors later left to found or join competing AI ventures. Treating today’s licensing-and-hiring deals as quasi-acquisitions of durable market power risks missing how AI knowledge and talent actually spread. 

The policy risk is asymmetric. Auer and Mario Zúñiga have warned that aggressive scrutiny of AI partnerships and acqui-hires may create the “very harms that policymakers wish to avert”: deterring incumbents from entering generative-AI markets through strategic relationships and making it harder for AI startups to obtain the funding they need to scale and compete. 

That concern is not theoretical. The deals CADE has reviewed funded the startups involved. Character.AI received roughly $2.7 billion through its Google deal; Inflection received roughly $650 million through its Microsoft deal. Jonathan Barnett similarly argues that preemptive antitrust fits poorly in the early-stage generative-AI ecosystem, where uncertainty is high and premature intervention can suppress efficient arrangements that combine the complementary assets needed to build AI models and applications. 

Taken together, these considerations counsel against an intensive enforcement focus on early-stage AI partnerships. A more measured skepticism toward antitrust intervention is the sounder posture, given the market dynamics and funding structures that characterize AI startups, and given the absence—at least in the current literature and available evidence—of a credible theory of harm for digital killer acquisitions. 

A Cautious Start, With One Big Exception

CADE’s recent decisions show that Brazil is entering a new phase of antitrust enforcement in AI markets. The authority has made clear that emerging partnerships between incumbents and startups deserve close scrutiny. That is a defensible position. Scrutiny, though, is not the same as a presumption of harm. 

To its credit, CADE took the prudent course by dismissing the Administrative Procedure for Investigation of a Concentration Act (APAC) proceedings in NVIDIA/Run, Microsoft/Mistral, and Google/Character.AI. The Tribunal also emphasized that the Article 88, §7 call-in power is an exceptional tool, to be used only when there is solid evidence of likely competitive harm in Brazil. That approach is consistent with what we have argued elsewhere: Brazil’s antitrust regime should resist the temptation to recreate the European Union’s Digital Markets Act (DMA) by other means. 

CADE did not exercise the same restraint in Microsoft/Inflection. The agency will now subject a transaction completed in 2024 to ex post merger review roughly two years after closing, even though the CMA cleared the same transaction after concluding there was no “realistic prospect of a substantial lessening of competition” on the same facts. In doing so, CADE elevated the digital killer-acquisition theory from an academic debate to the doctrinal foundation for intervention.

If the 2024 Joint Statement called for international convergence, the irony is hard to miss. The agencies championing that effort have thus far found no competitive harm in the AI partnerships they have reviewed. CADE reached that same conclusion in three cases. Microsoft/Inflection stands as the exception.

The story is far from over. Microsoft/Inflection will now receive a full substantive merger review. Amazon/Anthropic remains pending, with a decision expected after further inquiry into the “notorious facts” identified by Amaral. Google/Windsurf and Google/Hume AI are also making their way through the same APAC process.

CADE will have many more opportunities to shape its approach to AI partnerships. The hope is that it follows the evidence rather than the narrative. In markets that move this fast, antitrust should be careful not to mistake investment for foreclosure, or collaboration for consolidation.

The post Killer Theories and Acqui-Hire Alibis appeared first on Truth on the Market.

In Trump v. Slaughter, the Court overruled that 1935 opinion. The president may now dismiss the heads of federal executive agencies—including members of the Federal Trade Commission and certain other “independent agencies”—at will. So says the Supreme Court. 

Humphrey’s Executor was something of an odd duck even in 1935, on both constitutional and factual grounds, and the Court had been whittling away at its reach for decades. As Chief Justice John Roberts put it, writing for the majority: 

Humphrey’s is now far out of step with our cases, which have all but limited it to its facts.

And:

If anything more is left of Humphrey’s, we overrule it. Humphrey’s has for decades been a result in search of a rationale.

Or as the Python poets might have put it:

This precedent is no more! It has ceased to be! It’s expired and gone to meet its maker! This is a late case! It’s a stiff! Bereft of life, it rests in peace! If you hadn’t nailed it to the perch, it would be pushing up the daisies! It’s rung down the curtain and joined the choir invisible. This is an ex-precedent!

That seems a bit overenthusiastic. I have been in no way prescient in declaring the writing on the wall for Humphrey’s Executor. More than that, I believe the Court made the right decision.

And that is not because Rebecca Slaughter and I have very different approaches to antitrust law and economics, although we do. Night and day, really. Then again, I have some differences with everyone I know, including those who—grâce à Dieu—pay me to be myself, and that is no reason to reorganize a government. 

Rather, as I said before, her firing—the issue in the case—along with that of Alvaro Bedoya, which was a done deal either way, seemed to me:

on balance, a shame; that is, it seems to me unfortunate as a matter of practice, if not as a matter of political or constitutional principle. And for the most part (at least, for a long time), the odd statutory structure of the FTC seemed to work in practice. Indeed, it seemed to work pretty well.

There are always tradeoffs, and I don’t mean to pretend that nothing was lost, or put at risk, when the Court made the eminently justifiable decision that nearly all of us saw coming. For thoughtful, historically grounded consideration of those policy tradeoffs, I recommend an excellent article by William Kovacic, a former FTC chairman, and Marc Winerman, a former colleague of mine and longtime FTC staffer, “The Federal Trade Commission as an Independent Agency: Autonomy, Legitimacy, and Effectiveness.” 

Nor do I mean to suggest the decision was a constitutional or jurisprudential no-brainer. It seems clear to me that Chief Justice Roberts, writing for the majority, had the better argument, while Justice Sonia Sotomayor, writing in dissent, had the worse. But the questions were real. 

I have written at length about these issues elsewhere and won’t rehash them here—mostly, I commend the Court’s decision itself to those who want the details. Still, there were reasons to wonder about those details, and about the case’s implications. There is no doubt that Slaughter brought a colorable argument to the U.S. District Court for the District of Columbia when she first sought reinstatement. She had both the 1935 precedent—narrowed by then, but not yet overruled—and the plain statutory language of Section 1 of the Federal Trade Commission Act on her side. 

For all that qualification, I think the Court got it right. The alternative would have been something of a muddle, at best, as was Humphrey’s Executor itself. But more than that, my perhaps all-around contrarian take is this: Maybe this is less of a big deal than most observers, on either side of the decision, seem to think. 

The Beacons of Gondor Remain Unlit

The blogosphere is all atwitter over this one, as is the social-media platform formerly known as Twitter. No, it is not getting the same level of attention as the June 30 opinion in Trump v. Barbara, the birthright-citizenship case—and it shouldn’t. But it is no small matter, and there is no little drama afoot. 

That this is the first, second, or third place you have read about either decision seems unlikely. Here are the Wall Street Journal, the Washington Post, and the New York Times, to pick a few major outlets. I will pilfer a bit from each of my prior posts, but if you want to pick just one for review, I recommend the first: “Termination Tuesday: A Quasi-Comprehensive, Quasi-Definitive Discussion of the FTC and Humphrey’s Executor.” 

All that discussion notwithstanding, there are things to clear up. Intelligent people from other fields have asked me about this, including, e.g., about a political scientist who claimed that the decision will make it “SUBSTANTIALLY EASIER for Presidents to weaponize the state” (caps in original) and that “authoritarianism will follow.” That is, to be fair, only slightly more overdramatic than Justice Sonia Sotomayor’s dissenting statement, joined by Justices Elena Kagan and Ketanji Brown Jackson, that “the majority’s theory . . . promises to unleash only chaos.” 

One might well worry about presidents and lesser officials weaponizing the state. One might have such worries independent of partisan affiliation or one’s assessment of this administration, the one before it, or the 12-year term served by Franklin D. Roosevelt—who, of course, served before the adoption of the 22nd Amendment, and both before and after he proposed to pack the Court through the Judicial Procedures Reform Bill of 1937. 

One might worry more broadly about the politicization of law enforcement. While we are at it, and wondering about the Constitution’s separation and limitation of government powers, we might worry about the executive branch’s chronic arrogation of power since—well, at least since Thomas Jefferson. One might even worry about chaos, much as it seems to be the way of things.

But this decision does not herald doom. Nor does it end expert administration, such as it has been. 

Readers of Truth on the Market know full well that there are two federal antitrust agencies: the FTC and the Antitrust Division of the U.S. Department of Justice (DOJ). You also know that the head of the Antitrust Division—an assistant attorney general—serves at the pleasure of the president, as do the attorney general, the deputy attorney general, the solicitor general, the 94 U.S. attorneys, and others. 

The DOJ—more than 100,000 employees strong and our nation’s central law-enforcement agency—is not an independent agency. Neither Humphrey’s Executor nor Trump v. Slaughter had any bearing on the president’s ability to fire high-ranking DOJ officials at will. 

Remember Pam Bondi’s dismissal—sorry, I mean “transitioning to a much needed and important new job in the private sector”—as announced by presidential social-media post on April 2? A bit abrupt, that. But I do not recall anyone lighting the Beacons of Gondor to warn that civilization itself was threatened. How about President Donald Trump’s 2017 dismissal of acting Attorney General Sally Yates? There is a history of this sort of thing at DOJ, and not just across Trump administrations. 

Not incidentally, the DOJ—untouched by Trump v. Slaughter—has both civil and criminal enforcement authority, while the FTC has only civil authority. 

And, of course, there are the Department of Defense, with more than 600,000 civilian employees and more than 1.3 million active-duty troops; the Department of Homeland Security; the Department of the Treasury; the Department of State; the Department of Health and Human Services; the Department of Agriculture; and others. All are considerably larger than the FTC. And all, under settled constitutional doctrine, are headed by administrators who serve at the pleasure of the president. 

Moreover, while it is not exactly a done deal, there are reasons to think the Federal Reserve will remain untouched by the Court’s decision in Trump v. Slaughter. Not least among them is the Court’s decision in Trump v. Cook—which, like Trump v. Slaughter, was also handed down June 29. 

That case concerns the president’s August 2025 attempt to fire Lisa Cook, a member of the Fed’s Board of Governors and the first member any president had attempted to fire in the system’s 111-year history. The Court’s decision in Trump v. Cook was not a final decision on the merits. On the other hand, it denied the president’s application to stay a lower-court order rejecting the attempted dismissal, and it held that “[t]he Government has not shown that it is likely to prevail on the legal arguments advanced in its stay petition.” In doing so, the Court recited various grounds on which it might distinguish between the Fed and the FTC. 

Independence, With Scare Quotes

We might also wonder about the nature of the independence at issue with the FTC and other independent agencies that may mourn the passing of Humphrey’s Executor. As antitrust and administrative-law scholar Daniel Crane wrote in “Debunking Humphrey’s Executor”: 

At the end of a one- hundred-year natural experiment, the Commission bears almost no resemblance to the Progressive-technocratic vision articulated by the Court. The Commission is not politically independent, uniquely expert, or principally legislative or adjudicative. Rather, it is essentially a law enforcement agency beholden to the will of Congress.

In sum:

The upshot is that the FTC has essentially become the executive agency that the Humphrey’s Executor Court denied it was. The FTC functions primarily by enforcing the antitrust and consumer protection laws as a plaintiff, no more expert than the executive branch agencies doing the same thing. The principal structural difference from the executive branch agencies is that the FTC is beholden to Congress rather than to the President.

As I wrote earlier, a good many scholars agree—not all, but not a few. For dueling perspectives in the Harvard Law Review, compare Aditya Bamzai and Saikrishna Bangalore Prakash with Noah Rosenblum and Andrea Scoseria Katz. Moreover, empirical studies that examine independent agencies’ independence, and their expertise relative to other executive agencies, bear Crane out. If the question turns, in the end, on whether the FTC routinely carries out executive functions just as incontrovertibly executive agencies do—just as Cabinet-level agencies do—then Crane’s observation seems hard to contest. 

As for independence, many—including the dissent in Trump v. Slaughter—have noted that no independent agency is wholly independent. For more, I can once again recommend William Kovacic and Marc Winerman—both the article linked above and another, “The William Humphrey and Abram Myers Years: The FTC from 1925 to 1929.” 

Trivially, the FTC depends on Congress for funding, as well as for its many statutory charges. And all members of the bipartisan commission are, under the FTC Act, nominated by the president in the first place, subject to Senate confirmation. 

If there is an open “Democratic” seat on the commission—really, the statute just says that “[n]ot more than three of the Commissioners shall be members of the same political party”—when the president is a Republican, the nomination of a Democrat comes from that Republican president. Rebecca Slaughter, the respondent, was initially nominated by President Trump. And the president gets to decide—at the drop of a hat, with no further approval required—which sitting commissioner serves as chair, thereby investing that commissioner with substantial agenda-setting and administrative authority. The chair still gets just one vote on matters requiring a vote, but in all other ways, and by no small margin, the chair is first among equals. 

For that matter, I was at the FTC when Lina Khan served as chair, and when Rebecca Slaughter served as a majority commissioner, voting in lockstep with the chair. Both were duly nominated to the commission—Khan by President Joe Biden—and both were duly confirmed. Biden designated Khan as chair almost immediately upon her confirmation as a commissioner, as the statute permits, even if it came as something of a surprise. That was all well and proper, as far as that goes. 

Then again, those were years when the “independent” FTC seemed highly partisan—and partisan to a degree I had not seen at any time in my career. Both federal antitrust agencies touted a “whole of government” approach to enforcement that was not wholly unique, and that one might applaud, more or less, depending on the facts and circumstances. But for better or worse, it was hard to see any significant daylight among the White House, the “independent” commission, and the executive-branch Antitrust Division on matters of antitrust policy or administration. 

Uniformly, they sought to overturn established policy, including policy documents, and adopt new ones. Biden, Khan, and Jonathan Kanter, then the head of the Antitrust Division, all lamented what seemed to them decades of misguided and anemic antitrust enforcement, often in strikingly similar language. 

And as I noted in a prior post, I cannot help but recall the commission’s redaction—actual censorship—of then-Commissioner Christine Wilson’s dissent in the Meta/Within matter, or the much-reported staff gag order (see this staff report from the House Committee on Oversight and Accountability; or Politico, if you prefer). Departures of experienced senior staff were unusually numerous. 

There was not just some heavy-handed, top-down input on technical details of expert staff reports and investigations. I can recall at least two advanced-draft reports where staff were instructed not merely to table projects previously approved by leadership—to be fair, allocating staff resources was the chair’s prerogative—but to withhold embargoed drafts even internally. Those instructions expressly directed that the drafts be withheld from sitting FTC commissioners, such as Christine Wilson and Noah Phillips, who happened to be Republican members of the independent, bipartisan commission during a Democratic administration. 

For my own part, I had thought the four decades of antitrust enforcement lamented by Biden, Khan, and Kanter—and let’s not slight Tim Wu—were something much better, and indeed more effective, than either their lampoon or their alternative. Not always right, by any means, but often, and progressively so. 

When I joined the FTC in 2006, to work in an excellent Office of Policy Planning under the direction of Maureen Ohlhausen, the agency seemed to me something of a model for a research-based enforcement agency attentive to the risks of both false negatives and false positives. In antitrust, false negatives mean failing to challenge harmful conduct; false positives mean wrongly condemning conduct that benefits competition and consumers. Both types of error matter. 

Then again, there was only so much independence from elected officials even then. Appointees were, as always, appointees, subject to confirmation and funding. And while the FTC and DOJ did not march in lockstep from 1980 to 2020, they were in substantial agreement most of the time, on most issues, institutional differences and the occasional tiff notwithstanding. Perhaps the FTC leaned a bit more toward Congress and the DOJ a bit more toward the White House. But they did not disagree all that much, or all that often, on basic questions of merger-enforcement policy. That’s not to say they should have disagreed more. But it is to question whether the independent antitrust agency was systematically different from, much less better than, the dependent one.  

Who Are the Experts Again?

It also bears asking who, exactly, are the supposed experts in this expert agency. The question may seem impertinent. Slaughter and Bedoya, the two FTC commissioners fired by President Trump, are both intelligent, well-credentialed lawyers. So is Lina Khan. Under the law, and under executive and administrative tradition, all three were eminently qualified. 

But none came to the job with substantial administrative experience, or with much experience litigating matters on behalf of antitrust plaintiffs—inside or outside the government—or defendants. Khan was not even five years out of law school when she took the helm of the FTC, and Slaughter was not quite a decade out when she was nominated as a commissioner. None had a substantial background in relevant areas of economics, or any other areas of economics. And, pace Khan’s much-publicized student note in the Yale Law Journal, none had generated a substantial body of antitrust scholarship. 

If one likes expert agencies—and I do, at least in the right contexts and subject to appropriate constraints—one might still ask whether these are the sorts of experts who are supposed to be isolated from executive oversight so they might, as the Court writes, “deliver us to a promised land of technocratic governance.” 

We have had more experienced commissioners, to be sure. I have been privileged to serve with and under some excellent practitioners and scholars during both Democratic and Republican administrations. Truly. And, certainly, we have seen expertise and experience at the head of other agencies. 

But it is a crapshoot. Excellence in a relevant technical field is not now, nor has it ever been, required of presidential appointees. Not in this administration or any other; not at the FTC or any other federal agency. Neither is serious management experience. 

Consider a much larger expert agency under direct executive control: the U.S. Department of Health and Human Services. Within HHS, we find, among other subagencies, the Food and Drug Administration, the National Institutes of Health, and the Agency for Healthcare Research and Quality. The HHS secretary—Robert F. Kennedy Jr.—serves at the pleasure, or at least tolerance, of the president. Can anyone identify a relevant field in which Kennedy possesses relevant expertise or experience? Can anyone identify a historical administration wholly lacking in dubious appointments? 

Of course, there is no small amount of expertise at the FTC or, for that matter, at HHS. But the technical expertise lies mainly in the staff, not in the appointees. Neither Humphrey’s Executor nor Trump v. Slaughter addressed questions of staff tenure. 

I understand this might seem petty, churlish, or both—a lowly ex-staffer dissing his political-appointee betters—and I am sorry about that. Not one of those people ever did me any harm. Nor did I ever imagine that the prize of a presidential appointment was rightfully, or even reasonably, mine. Honestly, I think I would have been unsuitable—and not just unlikely—as an appointee under any administration. That’s on me. But it is neither here nor there. 

The ability of technical experts to bring their expertise to bear in service of the people of the United States always depends on appointed leadership, and on both Congress and the executive branch. That is equally true at the Securities and Exchange Commission, which is an independent agency, and the FDA, which is not. 

I am not saying that is a bad thing. Agents of the public should be publicly accountable, and executive agents especially should be accountable to the elected executive. But that arrangement provides only so much freedom from political input. Whether it works depends as much on the people involved as it does on institutional design and statutory removal protections. 

None of this endorses chronic or reckless interference in the details of law enforcement—civil or criminal, truly executive or quasi-executive. So, for example, I made the following two observations last September, after two senior Antitrust Division officials were reportedly “fired for insubordination”:

On the one hand:

As a formal matter, the U.S. attorney general [who answers to the President] is head of the department and does have administrative authority over the division. It’s not obvious that this authority should never be used.

On the other:

…there are at least strong pragmatic reasons to think that such authority should be used sparingly—indeed, rarely, when it comes to decisions about specific investigations, cases, and settlements. High-level policy may be the realm of the AG and, ultimately, the President, but high-level policy can only say just so much about the proper outcome of a matter subject to the rule of reason (something agency leadership under the Biden administration sometimes forgot).

Sea Change? You’re Going to Need a Smaller Boat

Perhaps the most interesting opinion of the day was Justice Neil Gorsuch’s concurrence, which raised serious but uneasy questions about the case’s implications. Little things: separation of powers, delegation and nondelegation, limited government, etc. To take just his opening: 

To fulfill his constitutional duty to ensure the laws are faithfully executed, the Court holds, the President must have the ability to remove principal officers who exercise executive power in his name. That includes those who run independent agencies like the Federal Trade Commission (FTC). With all this, I agree.

But neither can I ignore the implications that follow. Today, independent agencies do not just exercise executive law-enforcement powers. Congress has also delegated to them vast legislative and judicial powers, effectively allowing these agencies to make laws and decide disputes under them. And, after today’s decision, the President can effectively exercise all those powers too.

It’s a development that raises important questions, not least these: Would Congress have delegated so much power, including legislative and judicial power, to independent agencies had it known that the President would come to control them? How will Congress respond now—if realistically it can? And what, if anything, will this Court do about it?

Those are serious questions. About that, I think “this Court” will do rather less than some fear, and to unspectacular effect. But the details will matter all the same, and they remain to be seen. 

As for what difference this specific decision will make—well, something, and perhaps not a little. But if we are worried about preserving the independence of expert agencies, we might ask a further question about the degree of independence that was actually at stake: What the hell does it take to fire a commissioner for cause or, as the statute stipulates but does not define, “for inefficiency, neglect of duty, or malfeasance in office”? 

I am not suggesting that Slaughter was culpable under any of those standards. I am wondering about the substance of the standards themselves, which seem rather malleable as standards go and, for the most part, untested. Inefficiency? Really? Inefficiency, thy name is government. What would it take for a court to deny an executive determination that an appointee was inefficient? 

Consider the overruled case, Humphrey’s Executor, so named because William E. Humphrey had died by the time his case made it to the Supreme Court, which ruled in favor of his estate. Humphrey had been nominated to the FTC by Calvin Coolidge, our 30th president, and confirmed by the Senate. Concluding his initial term, he was reappointed by Herbert Hoover, our 31st president, in 1931. Hoover was succeeded by Franklin Delano Roosevelt, who did not much care for Humphrey and, in 1933, asked him to resign. Humphrey, caring little for resignation or Roosevelt, declined. 

And so, to quote the Supreme Court in Humphrey’s Executor, “on October 7, 1933, the President wrote him: ‘Effective as of this date, you are hereby removed from the office of Commissioner of the Federal Trade Commission.’” 

Roosevelt made no attempt to fire Humphrey for cause, just as Trump, more than 90 years later, made no attempt to fire Slaughter for cause. Neither president felt the need. What the Supreme Court decided in 1935—“distinguishing” (read “ignoring”) its own precedent in Myers v. United States—was not that Roosevelt had tried and failed adequately, or even nonpretextually, to establish “inefficiency, neglect of duty, or malfeasance in office.” It was that he had failed to so much as go through the motions. 

And when Trump dismissed Slaughter under the now-sustained banner of executive authority, it was not because he or his staff had tried and failed to make an adequate or nonpretextual case for inefficiency, neglect of duty, or malfeasance in office. It was because he, like Roosevelt, thought he should not have to bother. And because he wanted the Court to agree with him. 

And so it did.

Executive indifference to process—or even attempts to repudiate it—seems to be the way of things when the judiciary gets involved. 

But I rather suspect Slaughter could have been fired for cause had the administration wanted to bother, and not because of any malfeasance on her part. In any case, the confirmation of Mark Meador, who sailed through confirmation in April 2025, would have made her a minority commissioner, no more able to block a Republican FTC majority than Noah Phillips and Christine Wilson had been able to stop, slow, or substantially modify the commission’s partisan “reforms” under Lina Khan.

That is not to say minority dissents serve no function. Nor is it to deny the potential utility of bipartisan agreement—not just in cases where there is thoroughgoing bipartisan agreement, and there are such cases, but in cases and rulemakings where commissioners see both the potential for and utility of compromise in the service of unanimity. 

I have said it before, and I meant it: The underlying fracas seemed to me both unnecessary and a shame. 

Once more, with feeling: There are always tradeoffs. But not every tradeoff is a sea change.

The post Much Ado About Removal: The Supreme Court, the FTC, and the End of Independent-ish Agencies appeared first on Truth on the Market.

The Commission’s preliminary view that Amazon Web Services (AWS) and Microsoft Azure should be designated as Digital Markets Act (DMA) gatekeepers for cloud-computing services is more than another skirmish in Brussels’ long campaign against large technology companies. It is a test of whether the European Union will use competition policy to protect competitive rivalry or to administer digital markets by regulatory fiat.

The Commission says AWS and Azure are, respectively, the largest and second-largest cloud providers in the EU and that each serves as an “important gateway.” Yet no cloud provider satisfies the DMA’s ordinary quantitative thresholds. A formal decision is expected later in 2026, reportedly in December. The intervening months should not become a procedural pause. They should be used to ask the question the preliminary finding largely assumes away: Do enterprise cloud services fit the gatekeeper model at all?

The answer is no. The Commission’s cloud theory would extend an already defective ex ante regime into a market defined by enterprise procurement, multi-cloud strategies, rapid technological change, falling unit costs, and substantial ongoing investment. It also overlaps with the EU’s own cloud-specific Data Act, which already addresses switching, portability, interoperability, and egress fees. If the Commission’s goal is to make Europe more competitive and innovative, these proposed designations would move policy in precisely the wrong direction.

The Bureaucrat as Market Referee

The DMA was sold as a competition measure, but it is not competition law in the traditional sense. Classic antitrust, at its best, asks whether challenged conduct harms the competitive process, weighs efficiencies, and tailors remedies to proven harms. The DMA instead imposes categorical duties once a company and service receive the gatekeeper label. As I argued recently here at Truth on the Market, the regime replaces case-by-case proof of competitive harm with obligations triggered by administrative classification. That shift matters. It makes regulatory error cheap for the state and costly for markets.

The problem is not that all large platforms are harmless, or that policymakers should ignore exclusionary conduct. It is that ex ante intervention in dynamic markets carries high false-positive costs. Rules that ban or redesign integrated product features can destroy consumer benefits before anyone measures them. Access and interoperability mandates can reduce security, weaken investment incentives, or turn differentiated products into regulated utilities. Rules that force a firm to make rivals’ business models easier can protect competitors rather than competition. In digital markets, where value often comes from scale, integration, learning, reliability, and rapid iteration, these costs are not incidental. They are the main event.

Early DMA implementation already counsels caution. Some studies and market evidence report consumer frustration and business harm from compliance-driven product changes. The Nextrade Group survey of EU consumers and the European Centre for International Political Economy (ECIPE) study on consumer experience emphasize increased friction and limited visible benefits. In travel search, Mirai reported sharp declines in Google Hotel Ads clicks and bookings in DMA-affected markets, with intermediaries gaining visibility at the expense of direct hotel channels. These findings remain contested and do not settle the full welfare question. But they do puncture the tidy assumption that DMA compliance always ratchets toward consumer benefit.

For that reason, the EU should not expand the DMA into cloud unless the economic case is especially strong. The Commission’s preliminary case is not. It relies on size, investment capacity, AI partnerships, alleged lock-in, and cloud’s importance to Europe’s digital economy. Those facts may show commercial significance. They do not establish gatekeeping in the relevant economic sense.

Where’s the Gate?

The strongest objection to the cloud designation is conceptual. The DMA’s core platform-service framework was built for services that sit between businesses and consumers: search engines, app stores, social networks, operating systems, advertising services, and other digital chokepoints that can influence which firms reach end users. Cloud computing is different. It is business-to-business infrastructure. Cloud providers supply computing power, storage, databases, networking, security, and software tools. Their enterprise customers then use those resources to build products, applications, and internal systems. The cloud provider generally does not determine whether those customers reach end users, how users find them, or what prices they charge.

That is why International Center for Law & Economics (ICLE) Director of Competition Policy Dirk Auer argues that applying the DMA to AWS and Azure “stretches the act into a market it was never built to reach.” In its June 2026 assessment, ICLE noted that no cloud provider meets the DMA’s quantitative thresholds, that AWS and Azure hold estimated EU market shares of roughly 28% and 21%, respectively, and that Google Cloud accounts for about 14%. It also found that roughly 70% of customers use more than one cloud provider. Those are not the hallmarks of a tipped—or tipping—market. They are the hallmarks of vigorous competition among differentiated suppliers.

Lazar Radic makes the same point more fundamentally in his commentary “Cloudy Logic”: cloud has “no gate” in the relevant DMA sense. Enterprise customers routinely use multiple providers, negotiate contracts, benchmark performance, repatriate workloads, and combine hyperscale services with on-premises systems, European providers, telecommunications companies, specialized AI clouds, and open-source tools. A large bank, manufacturer, airline, or startup that relies on AWS for some workloads and Azure or Google Cloud for others bears little resemblance to a consumer locked into a single app store or search engine.

The Commission appears to treat switching costs as presumptive evidence of anticompetitive conduct. That is a mistake. In complex infrastructure markets, switching costs often reflect productive integration, customization, security choices, regulatory compliance, data architecture, employee training, and customer-specific investments. A sophisticated customer may deepen its relationship with one provider because it values reliability, speed, technical support, or specialized features. The fact that such a relationship is expensive to unwind does not prove exclusion. It may simply show that cloud services are differentiated products rather than interchangeable commodities.

Two Rulebooks, One Market

The overlap with the Data Act further weakens the Commission’s case. The EU’s own explanation of the law says Chapter VI is designed to make switching between data-processing services, including cloud and edge-computing services, “free, fast and fluid.” It requires providers to remove barriers to switching and multi-cloud use through measures such as open interfaces, machine-readable data exports, and functional-equivalence efforts for infrastructure services. It also abolishes switching charges, including data-egress fees, beginning Jan. 12, 2027.

That targeted framework already addresses the concerns driving the Commission’s cloud investigation: interoperability, portability, switching costs, and data access. Meanwhile, traditional EU competition law, particularly Article 102 of the Treaty on the Functioning of the European Union (TFEU), remains available to address specific exclusionary conduct by dominant firms. The added value of DMA designation is therefore difficult to see. The added costs are not: duplicative obligations, conflicting interpretations, compliance uncertainty, and litigation over which regime governs.

Regulatory overlap is not costless. Every new layer changes investment incentives. It encourages risk-averse compliance teams to slow product launches, narrow product offerings, or avoid features that could later be recast as self-preferencing, tying, or discriminatory access. It also rewards firms that excel at regulatory arbitrage rather than technical performance. Europe should want cloud customers to choose providers based on price, quality, latency, security, sustainability, and innovation—not on the unpredictable consequences of overlapping regulatory mandates.

Reading Draghi Backward

The proposed cloud designations are especially ill-timed in light of Mario Draghi’s report on EU competitiveness. Its central message is straightforward: Europe faces slowing productivity, intensifying global competition, and a widening innovation gap. The report argues that Europe has struggled to turn its scientific strengths into globally competitive technology companies and that fragmented and overly restrictive regulation has held back growth. The Commission has embraced Draghi’s agenda rhetorically. Expanding the DMA to cloud would undermine it in practice.

Cloud infrastructure is not a luxury for Europe’s innovation economy. It is the foundation on which businesses build artificial intelligence, cybersecurity, data analytics, biotechnology, advanced manufacturing, financial services, logistics, media, public-sector modernization, and startups. AWS and Azure are among the companies investing in that foundation. A regulatory regime that makes large-scale cloud investment less attractive, delays access to frontier AI services, or turns Europe into a compliance outlier will not create European champions. It will leave European businesses later to adopt new technologies, less productive, and less competitive.

A recent Amazon background paper makes the same point. It argues that European cloud customers enjoy broad choice, prices have fallen, and providers continue to invest billions in European data centers and related infrastructure. One need not accept every company claim at face value to recognize the underlying economics. Capital-intensive infrastructure markets depend on predictable returns. If Brussels signals that successful investment will trigger open-ended regulatory redesign, expected returns fall. Marginal projects get delayed, scaled back, or moved elsewhere. Compliance costs rise. Engineering talent shifts from building better products to building defensible ones. That is not a strategy for closing the technology gap with the United States or China.

The problem is not simply more regulation. It is regulation that treats scale as a liability. Draghi’s diagnosis points toward removing barriers to growth, investment, commercialization, and technological diffusion. Expanding the DMA points in the opposite direction. Europe cannot lament its shortage of globally competitive technology firms while treating scale, integration, and investment capacity as evidence of regulatory wrongdoing.

The Passport Test

A further concern is the DMA’s uneven application. The statute is formally nationality-neutral, but in practice, it falls overwhelmingly on large non-EU firms, particularly U.S. technology companies. The proposed cloud designations reinforce that perception. AWS and Azure are U.S.-based providers. Google Cloud, another U.S.-based hyperscaler, is the next-largest competitor. European providers, including OVHcloud, IONOS, Scaleway, telecommunications companies, and sovereign-cloud initiatives, remain outside comparable obligations.

There is no sound economic reason to treat a company’s nationality as a proxy for market failure. A U.S. cloud provider can benefit European customers, startups, public agencies, and exporters. A European intermediary can raise prices, reduce innovation, or lobby for protection. Competition policy should ask whether conduct harms the competitive process and consumers—not whether a successful company happens to be foreign.

The asymmetry also creates a political-economy problem. When regulation imposes costly obligations on a handful of foreign firms while redistributing traffic, access, data, or bargaining leverage to rivals and intermediaries, those beneficiaries gain a strong incentive to preserve and expand the regime. That is not competition policy. It is industrial policy dressed up in antitrust language.

Behind-the-Border Barriers, Brussels Edition

This is where Shanker Singham’s framework of anticompetitive market distortions (ACMDs), developed with colleagues including me, becomes useful. ACMDs are government-created or government-enabled measures that distort competition by raising rivals’ costs, restricting market entry, weakening property rights, or shifting economic rents to favored firms. They are “behind-the-border” barriers: not tariffs, but domestic rules that skew competition and reduce consumer welfare.

Viewed through that lens, the DMA—especially as applied to cloud—looks less like competition policy than an ACMD. It does more than prohibit proven exclusionary conduct. It imposes asymmetric obligations on designated firms because they are large and strategically important. It can require changes to product architecture, interoperability, data access, contracting, and business models. It shifts value from integrated platforms and their customers to rivals, intermediaries, and complainants who did not necessarily succeed through market competition. It also creates regulatory uncertainty that favors firms with the largest legal and compliance teams. In a cloud market that is already competitive and subject to sector-specific regulation, the DMA is more likely to distort competition than restore it.

The ACMD framework also exposes the limits of the Commission’s appeal to “fairness.” Fairness has no administrable limiting principle unless it is tied to consumer welfare, output, innovation, or a recognizable theory of competitive harm. A rule that makes life easier for competitors may make life worse for customers if it reduces quality, raises costs, delays new features, or discourages investment. A rule that fragments services may disadvantage users who value seamless integration. A rule that shifts rents from productive infrastructure providers to politically influential business users may harm the economy as a whole. Competition policy should not become a mechanism for politically allocating digital rents.

Calling It a Gatekeeper Doesn’t Make It One

The Commission should also pay close attention to the courts. In 2026, the General Court upheld parts of the Commission’s DMA designation of Meta but annulled the Facebook Marketplace designation because the Commission had not adequately justified its analysis. As I observed at the time, the ruling was narrow but significant. It underscored that labels, statutory categories, evidence, and reasoned decision-making still matter, even under an ex ante regulatory regime.

That lesson applies with even greater force to cloud. The Commission is not simply applying clear statutory thresholds to an obvious consumer platform. It is relying on a qualitative assessment to designate services that do not meet the DMA’s quantitative thresholds and that differ fundamentally from the services the statute was designed to regulate. If the courts insist on service-by-service analysis, current market evidence, and a genuine fit between legal categories and economic reality, the proposed cloud designations will face difficult questions.

The Commission could respond by issuing a longer decision. More pages would not fix the underlying problem. The real burden is not to produce a more elaborate explanation for forcing cloud into an ill-fitting category. It is to show that the designation will improve competition and consumer welfare after accounting for innovation costs, regulatory overlap, existing legal remedies, multi-cloud adoption, and the risk of deterring investment. On the current record, the Commission has not made that case.

Don’t Build the Gate

The Commission should change course before its final decision, expected in December 2026. It should not designate AWS as a DMA gatekeeper for cloud services. Nor should it designate Microsoft Azure. If a cloud provider engages in conduct that harms competition, the EU already has tools to address it: Article 102 TFEU, contract law, sector-specific rules, and the Data Act’s cloud-switching provisions. Those tools allow for a more tailored inquiry into conduct, effects, efficiencies, and remedies.

DMA designation would instead bring a blunt, overlapping, and innovation-hostile regime to a market that does not fit the gatekeeper model. It would compound the very weaknesses the Draghi report identifies: regulatory fragmentation, insufficient scale, weak investment incentives, and delayed adoption of frontier technologies. It would discriminate in practice against successful non-EU firms without establishing a sound economic reason to do so. And it would operate as an anticompetitive market distortion by burdening productive infrastructure providers and reallocating rents through regulation rather than competition.

Europe’s problem is not too much permissionless innovation from cloud providers. It is that too many promising technologies fail to scale, too much capital flows into compliance rather than growth, and too many regulatory initiatives treat market success as a public-policy defect. A stronger EU economy will not come from designating more gatekeepers. It will come from letting competition discover better products, lower prices, and more resilient infrastructure.

The Commission should withdraw its preliminary view, rely on the Data Act and ordinary competition law for specific cloud concerns, and resist turning the DMA into yet another anticompetitive market distortion. Cloud has no gate. Brussels should stop looking for one.

The post Brussels Goes Gate-Hunting: AWS, Azure, and the DMA’s Cloud Problem appeared first on Truth on the Market.

President Donald Trump announced June 10 that his administration is exploring taking equity stakes in major artificial intelligence companies, effectively making the federal government a part owner of the industry’s leading firms. The proposal echoes one made by Sen. Bernie Sanders (I-Vt.), who has argued that the American public should share in AI’s profits. Given that AI contributed an estimated full percentage point to real U.S. economic growth during the first three quarters of 2025 alone, the stakes are high.

But government ownership of AI would not merely “socialize the upside.” It would also socialize the downside, distort competition, expose taxpayers to political favoritism and corporate rescue missions, and give future administrations a powerful new lever over tools that mediate speech and information. The problem is not just industrial policy. It is industrial policy strapped to a First Amendment tripwire.

When the Referee Buys the Team

The administration has not specified how large those government stakes would be. If the model resembles the second Trump administration’s investments in at least 10 other companies, including Intel, the government would receive roughly 5%-10% equity stakes in exchange for billions in taxpayer dollars. Others have proposed going much further. Sanders and Steve Bannon, President Trump’s former chief strategist, have both suggested government ownership of up to 50% as a way to curb the power of what they call “tech oligarchs.”

The exact number matters less than the underlying principle. Even a nonvoting 5% stake creates distorted incentives that come with government ownership of strategically important industries.

Supporters argue that public ownership would let taxpayers share in AI’s success. It would also make taxpayers participants in its failures. Once the government becomes a shareholder, it gains strong incentives to shield those companies from failure, whether through additional taxpayer bailouts or preferential treatment in government contracting. That weakens the market discipline that encourages prudent risk-taking and continuous innovation.

The companies themselves would also face growing pressure to align with government priorities. Today, firms already risk losing contracts or attracting regulatory scrutiny when they clash with policymakers, as the recent Department of Defense dispute with Anthropic illustrates. Government ownership would amplify that leverage. Even without voting control, officials could threaten to sell a large stake, driving down a company’s stock price and giving Washington a powerful new tool to influence corporate behavior.

Government ownership would also tilt the competitive playing field. A disruptive startup that challenges a partially government-owned incumbent would no longer threaten only its rivals—it could also threaten the government’s own financial interests. That creates incentives to protect politically connected firms rather than reward the most innovative ones.

In effect, the government would be picking winners instead of allowing markets to discover them. The likely result: less competition, weaker incentives for new entrants, and reduced private investment—the very forces that have driven America’s leadership in high technology and artificial intelligence.

The arrangement also risks letting the tail wag the dog. Executives at partially government-owned firms would gain even greater leverage to lobby for favorable regulations or subsidies by arguing that policies harming their companies also harm taxpayers’ investments. It is hardly surprising, then, that some of the very “tech oligarchs” whom Sanders and Bannon criticize have expressed support for government ownership.

The economic risks are only part of the story. Government ownership would also inject politics directly into a technology that is becoming deeply embedded in everyday life. It would increase opportunities for censorship, propaganda, citizen surveillance, and “jawboning”—informal government pressure that coerces private companies into restricting or promoting speech.

History offers ample examples of governments using ownership or control of firms in critical technology and communications sectors to advance political objectives. Artificial intelligence could become the next arena where those temptations prove difficult to resist.

The Government Can’t Do by Proxy What It Can’t Do Directly

Freedom of speech is one of America’s defining constitutional protections. The First Amendment bars the government from directly censoring what people think, say, or post online. “Jawboning” refers to government efforts to sidestep those constitutional limits by pressuring private companies to suppress First Amendment-protected speech on the government’s behalf.

The Supreme Court recognized this principle decades ago in Bantam Books Inc. v. Sullivan (1963). There, Rhode Island officials asked book distributors for their “cooperation” in keeping certain publications off the market. The Court concluded that the state had devised an unconstitutional scheme to censor books indirectly, making clear that the government cannot evade the First Amendment simply by leaning on private intermediaries.

The Court reaffirmed that principle in National Rifle Association of America v. Vullo (2024), holding that government officials may violate the First Amendment by coercing private parties into punishing disfavored speakers. Those same constitutional principles apply to artificial intelligence, even if the technology presents the issue in a more complex form.

The Shareholder With a Badge

Large language models (LLMs)—the artificial intelligence systems behind tools such as ChatGPT, Claude, and Meta AI—are becoming a routine part of everyday life. Millions of Americans use them to write emails, summarize news, conduct research, generate computer code, and seek recommendations. AI does not deliver information the same way a search engine or social media platform does, but many of the same questions about information integrity remain. The sources an AI model relies on, the way it evaluates evidence, and the safeguards that shape its responses all influence what users ultimately see.

The First Amendment protects both the right to speak and the right to receive information. Government jawboning threatens both. It affects not only the person whose speech is suppressed, but also everyone who would otherwise hear it. The cases discussed above, along with more recent efforts by federal officials to pressure social media platforms to suppress discussion of COVID-19, illustrate that jawboning is often aimed less at silencing a particular speaker than at shaping what information reaches the public.

If the government directly instructed an AI company to alter its model by changing trusted sources, adjusting response filters, or otherwise steering its outputs toward favored viewpoints, the constitutional problem would be obvious. Government ownership presents a subtler, but potentially more powerful, mechanism for achieving the same result.

A casual suggestion from a regulator or a phone call from a senior official can carry enormous weight when the government also owns part of the company. Consider OpenAI. At a valuation of roughly $852 billion, a 5% government stake would be worth more than $42 billion. Even if those shares carried no voting rights, the mere threat that the government might sell such a large stake could send the company’s stock tumbling and pressure executives to accommodate official preferences.

The government’s leverage extends well beyond its equity stake. Unlike an ordinary shareholder, the federal government also controls taxes, procurement contracts, grants, export controls, and much of the regulatory environment in which AI companies operate. Faced with that combination of financial and regulatory power, executives would have powerful incentives to prioritize government approval over consumer welfare and innovation, or to accommodate whichever political or special-interest priorities happen to prevail at the time.

The Editor’s New Silent Partner

The Federal Trade Commission (FTC) recently sent Apple a letter alleging that its news aggregation service, Apple News, may violate the Federal Trade Commission Act by featuring fewer sources with a “perceived [conservative] ideological or conservative viewpoint” than those perceived as liberal. FTC Chairman Andrew Ferguson insists the agency is not policing speech, ideology, or editorial choices, which would raise serious First Amendment concerns. Instead, he argues that Apple may have misled consumers by creating the “reasonable expectation” that Apple News is an unbiased news aggregator.

Even assuming the FTC’s factual allegations are correct, that theory is difficult to square with the law. Private companies generally enjoy broad First Amendment protection to curate and moderate content. Apple News’ terms of use, which the FTC itself cites, expressly disclaim “ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING ANY WARRANTIES OF ACCURACY, NON-INFRINGEMENT, MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.” They also state that a user’s sole remedy for dissatisfaction with the service or its content is simply to stop using it.

Consumers have plenty of alternatives. Apple News competes with numerous news aggregators, each with its own editorial approach and recommendation algorithms. Consumers can also choose among news organizations with distinct editorial perspectives, just as they can watch Fox News instead of CNN. Both market themselves as independent news organizations, yet viewers understand—and often deliberately choose—them because of their perceived editorial slant. As Bill O’Reilly famously branded Fox News, it was the “No Spin Zone.”

Editorial judgment is not a market failure; it is one of the ways media companies differentiate themselves. Punishing firms for those choices would reduce, rather than expand, meaningful competition by making competing products look more alike. It would also sit uneasily alongside the Trump administration’s criticism of the European Union’s Digital Markets Act, which similarly subjects large, predominantly American technology companies to government scrutiny over how they curate and present content.

The same concerns apply to artificial intelligence. The FTC’s letter to Apple and the Federal Communications Commission’s (FCC) recent warnings that certain late-night television programming could face sanctions over allegedly biased political coverage illustrate how readily government officials can pressure private firms over editorial decisions that technically fall within existing legal authority. Giving the government an ownership stake in AI companies would hand future administrations an even more powerful lever.

That should concern supporters of every political persuasion. A Republican administration could use that leverage one way; a Democratic administration could use it another. Either way, government ownership would make it easier to pressure AI companies to suppress disfavored speech, promote preferred narratives, or tilt the competitive playing field in favor of politically connected firms.

Good Luck Proving It

Even when government pressure crosses the line into illegal jawboning, speakers and listeners can struggle to get relief. The problem is not always the principle. It is proof.

In Murthy v. Missouri (2024), several plaintiffs, including former Harvard professor Martin Kulldorff, sued the Biden administration for allegedly pressuring Twitter, YouTube, and other platforms to remove posts criticizing federal COVID-19 policies. Federal officials had, in fact, pressured the companies over such content. But a 6-3 Supreme Court majority sided with the government, holding that the plaintiffs lacked Article III standing because they had not shown a sufficiently direct causal link between government pressure and the platforms’ decisions to remove their posts.

That causation problem matters. A platform might have removed the speech under its own moderation policies, even if government pressure also played a role. Without proof that the government’s pressure caused a concrete and traceable injury, plaintiffs cannot sue. In practice, that often requires peering into the minds of platform employees—the sort of discovery courts tend not to hand out like Halloween candy.

Timing creates another obstacle. By the time a case reaches a judge, the officials involved may have left office, making an injunction against future jawboning harder to obtain or even moot. The result is a legal regime that recognizes the danger of informal government coercion but often struggles to deter it.

Those problems would only grow in the AI context. If officials pressured an AI company to alter how an LLM responds to politically sensitive questions, the company itself would often be the most directly injured party. But a partially government-owned company would have little incentive to sue the government that also sits on its cap table.

Market discipline can help—at least in a competitive market. If an AI model earns a reputation for biased, censored, or doctored responses, users can switch to rivals with better reputations. That threat gives companies a reason to protect the integrity of their outputs.

Government ownership weakens that discipline. Firms with a special relationship to Washington would enjoy advantages unavailable to actual or would-be rivals, making their market positions harder to challenge. That would deter entry, reduce competition, and make it harder for consumers to punish politically compromised AI tools by taking their business elsewhere.

The Best Fix Is Not Breaking It

The recently introduced Justice Against Weaponized Bureaucratic Overreach to Networked Expression (JAWBONE) Act, sponsored by Sens. Ron Wyden (D-Ore.) and Ted Cruz (R-Texas), reflects growing bipartisan concern that government pressure on private companies threatens Americans’ free speech. It also recognizes the practical reality that people often struggle to prove jawboning or obtain meaningful relief after it occurs.

The bill would create a private right of action, allowing individuals to sue government officials who coerce broadcasters, online platforms, or artificial intelligence companies into suppressing or manipulating lawful speech. Prevailing plaintiffs could recover damages and attorneys’ fees. The legislation also includes important safeguards. Communications made by government officials during lawful investigations, pursuant to warrants, or in the course of managing official government accounts would remain permissible.

Those reforms would meaningfully strengthen deterrence. By creating an explicit cause of action, the bill would help plaintiffs overcome some of the standing and causation obstacles that proved decisive in cases such as Murthy v. Missouri. It would also require the National Institute of Standards and Technology and the White House Office of Science and Technology Policy to establish a transparency regime documenting and publicly summarizing government communications with private companies about speech and content moderation. That recordkeeping would make it easier to identify and prove improper government pressure.

The JAWBONE Act strikes a sensible balance. It targets genuine coercion without preventing government officials from sharing important information with private companies during emergencies or other legitimate government activities. The bill would also apply to artificial intelligence companies, including those in which the government owns an equity stake.

Even so, legislation can address only part of the problem. Congress may amend the bill before passage, and there is no guarantee it will become law at all. More fundamentally, no statutory safeguard can fully eliminate the subtle pressures that arise when the government is both regulator and shareholder. The surest way to prevent officials from quietly steering AI companies toward preferred narratives is not to put the government on the cap table in the first place. Doing so would also avoid the broader costs to competition, innovation, and taxpayers that accompany government ownership or undue influence over a dynamic private industry.

Don’t Nationalize the Answer Machine

Government ownership of AI companies would not simply let taxpayers share in the industry’s upside. It would also socialize the downside, exposing taxpayers to losses while giving future administrations a powerful new way to pressure companies that increasingly shape how Americans seek, receive, and understand information.

That makes government influence over LLM outputs more than a tech-policy concern. It is a First Amendment concern, a competition concern, and a taxpayer concern rolled into one very bad bargain.

The government should not buy a seat at the table where Americans’ answers are written.

The post Don’t Put the Government on the Cap Table appeared first on Truth on the Market.

Also known as the App Store Accountability Act, SB 2420 is currently facing a major constitutional challenge before the 5th U.S. Circuit Court of Appeals. The consolidated cases—Students Engaged in Advancing Texas (SEAT) v. Paxton and Computer & Communications Industry Association (CCIA) v. Paxton—pit challengers against Texas Attorney General Ken Paxton and place the intersection of free speech and government regulation of technology platforms squarely before the court. 

SB 2420 requires app stores to verify the age of every user and mandates that minors obtain individualized parental consent before downloading or purchasing any app.

Texas argues that the law merely strengthens parental authority. The U.S. District Court for the Western District of Texas was unconvinced. As the district court explained: 

The Act is akin to a law that would require every bookstore to verify the age of every customer at the door, and for minors, require parental consent before the child or teen could enter and again when they try to purchase a book.

The case is now on appeal before the 5th Circuit. The International Center for Law & Economics (ICLE) filed an amicus brief supporting the plaintiffs and arguing that SB 2420 violates the First Amendment. The brief’s distinctive contribution is to connect First Amendment doctrine with an underlying law & economics framework, building on my prior ICLE issue brief, “A Coasean Analysis of Online Age-Verification and Parental-Consent Regimes,” as well as several earlier Truth on the Market posts examining app-store age-verification requirements. 

The Wrong Gatekeeper

ICLE’s brief centers on the concept of the “least-cost avoider.” In law & economics, the principle holds that liability should generally fall on the party best positioned to prevent harm at the lowest overall social cost. As I have noted elsewhere: 

From the perspective of law & economics, the most important question in disputes such as these is to identify the lowest-cost avoider of harms, while bearing in mind relevant transaction costs. It is not necessarily the case that the answer is always the user. It could be that online intermediaries are best-positioned to monitor and control harms to those users. But among the relevant social costs we must consider in this example is the risk of collateral censorship.

That risk is central to SB 2420. When the law makes app stores responsible for preventing potential harms to minors, it creates powerful incentives to overrestrict access to lawful content. To avoid liability, app stores must erect broad age-verification and parental-consent barriers around vast amounts of harmless, educational, and expressive material. The result is collateral censorship: protected speech becomes harder to access, not because it is harmful, but because platforms seek to avoid legal risk. 

App stores are poorly positioned to make these judgments. Requiring them to police every download and purchase imposes substantial economic friction while restricting access to First Amendment-protected expression.

Under SB 2420, minors cannot access a large swath of the modern marketplace of ideas without first obtaining a government-mandated digital permission slip. A minor could be blocked from freely accessing: 

• Current events through The New York Times app;
• Local weather information through The Weather Channel app;
• Vocabulary definitions through the Merriam-Webster Dictionary app;
• Scripture through the YouVersion Bible app; or
• Educational and creative content on platforms such as YouTube and Instagram.

Yet the Supreme Court has repeatedly recognized that minors enjoy substantial First Amendment protections. The government lacks a “free-floating power” to dictate which ideas children may encounter. SB 2420 nevertheless places state-mandated barriers between minors and a wide range of constitutionally protected speech. 

Parents and minors, by contrast, are often the true least-cost avoiders. They already have access to voluntary tools—including content filters, parental controls, and built-in app-blocking features—that allow families to tailor restrictions to their own preferences and circumstances. That reality also aligns with a core principle of First Amendment doctrine: when less restrictive alternatives exist, the government cannot justify burdening protected speech with broad, one-size-fits-all regulations. 

Calling Speech a Contract Doesn’t Make It One

On appeal, Texas has tried to minimize the First Amendment issues by recasting SB 2420 as a mere contract regulation. Because app stores require users to accept terms of service and often monetize user data, the state argues that it may freely regulate a minor’s ability to enter those arrangements. 

The 5th Circuit accepted a version of that argument when it stayed the district court’s preliminary injunction, treating the law as a regulation of “commercial speech” and subjecting it to the less demanding standard of intermediate scrutiny. 

Federal courts around the country, however, have repeatedly rejected similar arguments. As ICLE’s brief explains, there are at least two reasons why. 

First, speech does not lose First Amendment protection simply because someone earns money from it. A bookstore does not forfeit its constitutional rights because it sells books for profit, and an app store is no different. The Supreme Court has long recognized that entities engaged in expressive activity retain First Amendment protection even when they operate as businesses. 

Second, any commercial aspects of app-store transactions are “inextricably intertwined” with fully protected speech. App stores are multi-sided platforms that connect developers and users. Many apps are offered for free or at reduced cost because developers and platforms rely on advertising revenue generated through user data. The economic transaction and the dissemination of speech are therefore inseparable. 

For that reason, SB 2420 cannot be treated as a simple commercial regulation. Because the law restricts access to fully protected speech based on content, the district court correctly concluded that it must satisfy strict scrutiny—the most demanding standard of constitutional review. 

Parental Controls, Not State Controls

To survive strict scrutiny, a law must be the least restrictive means of advancing a compelling government interest. Even assuming the state has a compelling interest in protecting minors from potentially harmful content, SB 2420 fails that test because it is both overinclusive and underinclusive. 

The law is overinclusive because it restricts access to vast amounts of harmless, First Amendment-protected speech. It also forces minors—and adults who value their privacy—to surrender sensitive age-verification information merely to use an app store. For many apps, parents would likely have no objection to a download. Yet in a world with real transaction costs, requiring parental approval for every app and purchase means some speech will inevitably be restricted regardless of parental preferences. 

At the same time, the law is underinclusive. A minor can bypass the app store entirely by opening a pre-installed browser such as Safari or Chrome and accessing the same allegedly harmful content directly on the web. 

SB 2420 fares no better under intermediate scrutiny. The law burdens substantially more speech than necessary to achieve its stated objectives. There is no sound basis for restricting access to speech across every app simply because a small subset of apps may contain objectionable content. 

Helping parents protect their children online is a legitimate and worthwhile goal. The problem is that Texas chose one of the most speech-restrictive ways to pursue it. Parents already have access to voluntary tools, including content filters, parental controls, and app blockers. Market demand has given app stores strong incentives to provide those tools, and families remain free to use them as they see fit. 

The 5th Circuit should affirm the preliminary injunction. The First Amendment problem with SB 2420 is not that it seeks to protect children. It is that they try to accomplish that goal using the wrong gatekeeper. The least-cost avoiders are parents and minors—not the app stores that stand between them and the modern marketplace of ideas.

The post Texas Wants to Check Your App Store Papers appeared first on Truth on the Market.

The Act was designed to speed wireless deployment while preserving local control over routine land-use decisions. Section 332(c)(7) reflects that compromise. It preserves state and local authority over the “placement, construction, and modification” of wireless facilities, while imposing a handful of federal constraints to ensure Americans receive the benefits of timely wireless-service deployment. 

The most important of those constraints is the effective-prohibition clause, which provides that local regulation “shall not prohibit or have the effect of prohibiting the provision of personal wireless services.” The key words are “or have the effect of prohibiting.” That language extends beyond outright bans to government actions that, whatever their form, leave an area without wireless service. For nearly 30 years, courts have wrestled with a recurring question: How far does that functional phrase reach? 

Faced with a statute that condemned effects without defining them, the federal courts of appeals developed a framework to fill the gap. Beginning with the 2nd U.S. Circuit Court of Appeals in Sprint Spectrum, L.P. v. Willoth (1999), and eventually adopted by nearly every circuit to consider the issue, courts converged on the “significant gap” test. Under that approach, a denial has the effect of prohibiting service when it leaves a significant gap in a carrier’s coverage and the carrier’s proposal is the least intrusive means of closing it. The test gave concrete meaning to the statute’s “effect of” language, tied liability to real-world coverage rather than the label a locality attached to its decision, and used the no-alternatives requirement to supply the causal connection implied by the word “effect.” 

Last month, the 11th U.S. Circuit Court of Appeals broke from that consensus. In T-Mobile South, LLC v. City of Roswell, the court held that the effective-prohibition clause governs only the regulation of siting—that is, control through generally applicable rules—and therefore cannot be invoked to challenge the denial of a single permit application. 

That reading is difficult to square with the statutory text. The phrase “effect of prohibiting” is at least as naturally read to reach functional prohibitions as formal ones. The court’s narrower interpretation also carries consequences that cut against the deployment Congress sought to accelerate. Under the 11th Circuit’s approach, a locality can keep wireless facilities out indefinitely by denying applications one at a time, each on seemingly site-specific grounds, without ever adopting a rule that a court could invalidate. The result is a moratorium in all but name—effectively insulated from challenge because no one put it in writing. 

The Consensus Nobody Argued About

For nearly 30 years, most federal courts of appeals operated on a shared—and largely unexamined—premise: a city’s denial of a single wireless-facility permit can “have the effect of prohibiting” service under 47 U.S.C. § 332(c)(7)(B)(i). The debate was rarely whether an individual denial could violate the statute. The real question was how courts should determine when it does. 

The framework emerged almost immediately after enactment of the Telecommunications Act of 1996. In Sprint Spectrum, L.P. v. Willoth (1999), the 2nd U.S. Circuit Court of Appeals held that a denial effectively prohibits service when it prevents a carrier from remedying a significant gap in coverage and the proposed facility is the least intrusive means of closing that gap. 

Just as important, the court rejected the argument that Section 332 reaches only general bans on wireless service: 

Absent an explicit policy banning personal wireless services, the Board contends, courts can only consider whether in aggregate a town’s repeated denials of applications have the effect of a general ban. Since Ontario does not have a general ban in effect, as evidenced by its earlier approval of Frontier’s application and professed willingness to accept some level of PCS service, the Board insists its actions must necessarily be in conformance with subsection B(i)(II). We disagree with this reasoning.

Not every circuit followed the same path, but most arrived at the same destination. The 4th U.S. Circuit Court of Appeals initially suggested that the effective-prohibition clause targeted only general bans. See AT&T Wireless PCS, Inc. v. City Council of Virginia Beach (1998). Yet the court soon recognized that an individual permit denial could itself amount to an effective prohibition when it prevents a carrier from remedying a significant coverage gap and no reasonable alternatives exist. See 360 Communications Co. of Charlottesville v. Board of Supervisors of Albemarle County (2000). The court reaffirmed that understanding in T-Mobile Northeast LLC v. Fairfax County Board of Supervisors (2012), while emphasizing that the statute does not guarantee carriers flawless coverage. The burden may vary across jurisdictions, but the basic premise remained the same: individual siting decisions can be challenged under Section 332. 

Stripped to its essentials, the resulting “significant gap” test asks two questions. First, does a significant gap in wireless coverage exist in the relevant area? Second, do reasonable alternatives to the carrier’s proposed facility mean that service can be provided without the denied application? That two-step inquiry became the dominant framework for effective-prohibition claims nationwide. 

Over the next decade, most circuits adopted some version of the significant-gap approach. The 7th U.S. Circuit Court of Appeals did so in VoiceStream Minneapolis, Inc. v. St. Croix County (2003); the 9th U.S. Circuit Court of Appeals in MetroPCS, Inc. v. City & County of San Francisco (2005); and the 6th U.S. Circuit Court of Appeals in T-Mobile Central, LLC v. Charter Township of West Bloomfield (2012). 

If anything, more recent developments expanded the reach of the effective-prohibition clause rather than narrowing it. In a 2018 declaratory ruling, the Federal Communications Commission (FCC) recast the inquiry around what it called “material inhibition.” Under that standard, a state or local requirement effectively prohibits service whenever it materially inhibits a provider’s ability to offer a covered service. The FCC explained that the standard extends beyond coverage gaps to include barriers to network densification, capacity upgrades, and deployment of new technologies such as 5G. A locality need not erect an insurmountable obstacle; a material one is enough. 

The FCC derived the test from its 1997 California Payphone decision and applied it to the parallel effective-prohibition language found in both Sections 253 and 332. Although the ruling focused primarily on broad regulatory barriers, the 3rd U.S. Circuit Court of Appeals applied the standard to an individual permit denial in Cellco Partnership v. White Deer Township Zoning Hearing Board (2023). The court expressly abandoned its earlier significant-gap framework and instead asked whether, under the totality of the circumstances, the denial prevented the carrier from providing service without unreasonable cost. Where the significant-gap cases at least attempted to cabin liability, the material-inhibition standard lowered the bar further. 

The 11th Circuit Changes the Question

Last month, the Eleventh Circuit decided that the widespread application of section 332 to individual siting decisions was incorrect. Instead,  the court held that the statute’s bar on conduct that has “the effect of prohibiting” wireless service does not reach an individual permit denial at all. It limits only “[t]he regulation of” siting — and “regulation,” the court concluded, means control by rule, not the disposition of a single application.

Last month, the 11th U.S. Circuit concluded that nearly three decades of effective-prohibition jurisprudence had gone off track. In Roswell, the court held that Section 332’s prohibition on conduct that has “the effect of prohibiting” wireless service does not reach the denial of an individual permit application. Instead, it applies only to “[t]he regulation of” wireless-facility siting. And, in the court’s view, “regulation” means control by rule, not the disposition of a single application. 

The court’s reasoning is relentlessly textual. Because “regulation” derives from the verb “regulate,” the court began with contemporaneous dictionary definitions, which generally described regulation as controlling or directing conduct “according to rule, principle, or law.” A locality that denies a permit, the court reasoned, is not regulating property through a rule. It is merely refusing to grant one applicant an exception to rules already in place. The actual constraints on where a wireless facility may be built come from the ordinance itself, not from any individual permitting decision. 

The court found further support in Section 332’s structure. Subsection (c)(7)(A) preserves local “authority . . . over decisions” regarding wireless-facility siting. The limitations that follow in subsection (c)(7)(B), meanwhile, divide neatly into substantive and procedural categories. 

The substantive limits—the prohibitions on unreasonable discrimination and effective prohibition in subsection (B)(i), as well as the radiofrequency-emissions provision in subsection (B)(iv)—speak in terms of “regulation.” The procedural limits—the requirement to act within a reasonable time in subsection (B)(ii) and the requirement that denials be supported by substantial written evidence in subsection (B)(iii)—speak instead of a “decision” to “deny a request.” 

For the court, that distinction mattered. Congress had available a term broad enough to encompass both generally applicable rules and individual permit denials: “decisions.” Yet it chose “regulation” for the statute’s substantive constraints. When Congress uses materially different terms, courts generally presume it intends materially different meanings. On that logic, “regulation” is a subset of “decisions,” not a synonym for them. 

The implications are sweeping. The court held that the significant-gap test “fails at the threshold” because it focuses on the wrong object. The relevant question is not whether a permit denial leaves a coverage gap, but whether a governing rule has the effect of prohibiting service. The panel was openly skeptical of the significant-gap framework’s pedigree, describing it as the product of judicial “common-law rulemaking” untethered from statutory text and suggesting that its widespread adoption was evidence that “something has gone badly wrong.” 

Although the court framed its criticism in terms of the significant-gap test, its reasoning extends much further. The FCC’s material-inhibition standard likewise asks whether government action functionally impedes deployment. If the statute reaches only rules and not individual denials, that approach appears vulnerable for the same reason. As the court put it, its interpretation is “irreconcilable with any version of the significant gap test.” 

Still, the court did not eliminate effective-prohibition claims altogether. Providers may continue to challenge local regulations that operate as barriers to wireless deployment. A carrier that can show ostensibly permissive siting rules function as a de facto ban, or that a consistent pattern of denials reflects an unwritten policy against wireless facilities, may still prevail. What the carrier may not do, under the 11th Circuit’s approach, is treat the denial of a single permit application as the prohibited act itself. 

The Evasion Problem

The significance of Roswell lies not only in its sharp break from existing precedent, but also in how that break cuts against the purposes of the Telecommunications Act. If the decision stands, its effects may extend well beyond the 11th Circuit, creating new obstacles to wireless deployment and increasing uncertainty for providers nationwide. 

As a textual matter, the court’s critique of the significant-gap test is less persuasive than it first appears. Congress did not merely prohibit regulations that “prohibit” wireless service. It prohibited regulations that “prohibit or have the effect of prohibiting” the provision of wireless service. An outright ban already “prohibits.” The additional phrase must therefore do some independent work. 

The most natural reading is that Congress sought to reach government actions that, whatever their form, make service unavailable in practice. Viewed through that lens, the significant-gap test is not free-floating common law. It is an attempt to give operational meaning to the statute’s “effect of prohibiting” language. The test’s significant-gap requirement tracks the statutory concern with the provision of service, while its no-reasonable-alternatives requirement supplies the causal connection implied by the word “effect.” Without the alternatives inquiry, a court cannot determine whether a denial actually caused a service gap or merely foreclosed one of several viable paths to coverage. 

Indeed, even the 4th U.S. Circuit’s decisions suggest that the significant-gap framework can be used to determine whether an ostensibly individual siting decision is functionally equivalent to a broader policy of exclusion. Under that approach, a permit denial can violate Section 332 when it effectively operates as a prohibition, regardless of whether the locality formally labels it as such. In that respect, the significant-gap test is not obviously inconsistent with the logic that Roswell claims to embrace. 

The practical consequences of the decision are more troubling. Under Roswell, a single denial can never establish that a locality has effectively prohibited wireless deployment. Instead, a provider must identify a written or unwritten rule that forecloses service. That approach creates an obvious avenue for evasion. 

A locality determined to keep wireless facilities out need not enact a moratorium or adopt an exclusionary ordinance. It can simply deny applications one by one, each supported by seemingly site-specific findings about aesthetics, neighborhood character, or the particular parcel at issue. So long as no formal rule emerges, the locality may achieve the same result while avoiding scrutiny under the effective-prohibition clause. The result is a moratorium in all but name. 

That concern is hardly hypothetical. In its 2018 Small Cell Order, the FCC concluded that de facto moratoria—local actions that effectively halt deployment without formally banning it—can themselves constitute effective prohibitions. Roswell points in the opposite direction. Rather than looking at whether deployment has been blocked in practice, it rewards localities that accomplish the same objective through a series of individualized denials.

The panel acknowledged this concern but offered only a limited response. Providers, the court explained, remain free to challenge unwritten rules or to show that ostensibly permissive ordinances are merely a “fig leaf” for a prohibition. Yet the court also described that burden as “heavy.” The problem is that proving an unwritten policy is most difficult when intervention is most valuable: early in the process, before enough denials have accumulated to reveal a pattern. By the time a provider can demonstrate a systematic practice of exclusion, deployment may already have been delayed for years.  

The court’s fallback answer is that Congress remains free to amend the statute. But that response assumes the statute does not already address the problem. A stronger reading is that Congress addressed it directly. The phrase “have the effect of prohibiting” instructs courts to look beyond form and examine function. The significant-gap test, whatever its imperfections, represented a serious effort to do exactly that. Roswell replaces that inquiry with one focused largely on whether the locality had the foresight to avoid putting its prohibition in writing. 

The decision may also inject uncertainty into wireless deployment far beyond the 11th Circuit. The court did not reject the significant-gap framework on its own terms. Instead, it held that providers must first show a broader regulatory policy before the framework becomes relevant. Other courts could adopt that threshold requirement without formally repudiating their existing precedents. 

That uncertainty matters. Wireless infrastructure projects require substantial upfront investment, and providers must assess regulatory risk before committing capital. If carriers can no longer assume that an unlawful denial can be challenged promptly under Section 332, some projects—particularly those with more modest expected returns—may never move forward. The result would be slower deployment, weaker coverage, and fewer options for consumers at a time when demand for wireless capacity continues to grow. 

How to Ban Towers Without Banning Towers

Roswell marks a sharp break from nearly 30 years of Telecommunications Act jurisprudence. By insisting that only a “rule” can violate the effective-prohibition clause, the 11th U.S. Circuit removed the most common effective-prohibition claim from federal court and handed localities a roadmap for resisting wireless deployment without adopting a policy providers can challenge directly. 

That result is hard to square with the Act’s text or purpose. Congress barred not only regulations that prohibit wireless service, but also those that “have the effect of prohibiting” it. The significant-gap test was imperfect, but it took that command seriously by asking whether local action actually prevented service from being provided. 

Roswell trades that functional inquiry for formalism. Under its logic, the problem is not whether a locality blocks deployment, but whether it does so through a rule rather than one denial at a time. 

Congress enacted Section 332 to keep local authority from becoming a bottleneck to wireless deployment. The 11th Circuit risks turning that bottleneck into a blueprint.

The post The Roswell Loophole: How to Stop Wireless Deployment One Permit at a Time appeared first on Truth on the Market.

Sen. Chuck Grassley (R-Iowa) and Sen. Amy Klobuchar (D-Minn.) “introduced” AICOA last week. Co-sponsors include Sens. Dick Durbin (D-Ill.), Josh Hawley (R-Mo.), Sheldon Whitehouse (D-R.I.), and Cory Booker (D-N.J.). Here is the bill’s text, at least as introduced. 

“Introduced” is technically correct as a matter of process. And, as far as I know, this precise text string has not been introduced before. But we have seen pretty darn similar bills, under the same title, from Klobuchar before. AICOA appeared as S. 2992 in the 117th Congress (we’ll call that AICOA 1.0) and S. 2033 in the 118th Congress (AICOA 1.1). There have been changes along the way, but the essentially bad idea remains, in essence, bad. 

I don’t know whether AICOA 1.2’s bite at the apple will lead to anything more than the others did. There’s rather a lot going on, and I think I read something about an election to be held in November of this year. My best guess is that this version, too, will not pass. 

Then again, I lack a crystal ball. And there is bipartisan interest—not least among populists on the right and the left—in doing something to the sector. So I’m less sanguine about all this than I’d like to be. Is the third time the curse? 

Same Same But Different, And Not Better 

My International Center for Law & Economics (ICLE) colleague Geoffrey Manne has a helpful initial summary under the equally helpful title, “Revised AICOA Doubles Down on Flawed Antitrust Shortcut.” Manne’s top-level observation is this: 

The latest version of [AICOA] fails to fix the bill’s central legal and economic flaws—and in several ways makes them worse. 

That seems right, as do Manne’s more specific objections:

AICOA’s economic assumptions remain flawed. The bill treats vertical integration, self-preferencing, and default settings as suspect, even though those practices often make products better, safer, and easier to use. …

It lowers the competition-harm threshold to anything above de minimis. …

[It] raises the burden for key safety, privacy, and security defenses. …

The bill also replaces the earlier enforcement-guidelines process with expedited litigation provisions that direct courts to prioritize these cases and seek final judgment within one year. …

While the bill lowers the maximum civil penalty from 15% to 10% of U.S. revenue during the violation period, it adds a 1% floor once penalties are imposed. That is not leniency . . . It is a new floor for enormous fines, even in close cases involving unsettled legal questions. …

To that I’d add, among other things, a basic “big is bad” assumption lifted straight from the neo-Brandeisian playbook. See, for example, Tim Wu’s “The Curse of Bigness” and, for one possible if counterproductive implementation, the European Union’s Digital Markets Act. 

Analyses of AICOA 1.0 and AICOA 1.1 remain unfortunately relevant. Geoff’s list of prior ICLE scholarship is therefore useful as a current resource:

• “10 Things the American Innovation and Choice Online Act Gets Wrong” — Dirk Auer (Jan. 18, 2022)
• “How Tech Startups Could Be a Casualty of the War on Self-Preferencing” — Geoffrey A. Manne (Jun. 3, 2022)
• “AICOA Is Neither Urgently Needed Nor Good” — Thomas A. Lambert (July 25, 2022)
• “The Catch-22 of AICOA’s Guidelines” — Geoffrey A. Manne (July 25, 2022)
• “Reining in Digital Competition to No Good End: Will AICOA and OAMA Rise from the Grave?” — Daniel J. Gilman and Lazar Radic (Mar. 8, 2023) 
• “Regulate for What? A Closer Look at the Rationale and Goals of Digital Competition Regulations” — Lazar Radic, Geoffrey A. Manne, and Dirk Auer (Apr. 1, 2025) 

Many of our concerns were widely shared. Reviewing AICOA 1.0—the version from the 117th Congress—in the Michigan Technology Law Review, Herbert Hovenkamp summarized his view neatly: “AICOA was a bill that deserved to die.” 

His concerns were many, and familiar. Noting AICOA’s general hostility to self-preferencing—when a firm favors its own products or services on its own platform—Hovenkamp observed that “[s]elf-preferencing is an essential tool of competition, which has never imposed a requirement that people must sell other people’s merchandise.” He also objected that AICOA’s prohibitions applied “to products and services over which the seller has little or no market power. As a result, its substantive requirements are egregiously mistargeted.” 

Hovenkamp also anticipated that AICOA’s “gatekeeper” approach to competition policy would resurface—as indeed it has—despite the bill’s failure to gain traction in the 2021-2022 session: 

The issue … will almost certainly be considered again by Congress. When it does so, the “gatekeeper” approach to competition policy should be abandoned. It is too narrow because it ignores the conduct of firms that are not designated as gatekeepers, including offline sellers who are not included no matter what their size. It is too broad because it overreaches, perhaps egregiously, to condemn competitively harmless conduct by firms defined as gatekeepers.

That was hardly a defense of the antitrust status quo. To the contrary, Hovenkamp suggested that “[u]nderenforcement is a serious problem.” Set aside debates about the degree or locus of such problems. Hovenkamp recognized AICOA’s fundamental flaws as a vehicle for competition-policy reform—one with consequences likely, in some cases, to be unpredictable and, in others, “just plain bad.” 

Similarly, New York University’s Daniel Francis presented testimony on AICOA 1.1 before the Senate Judiciary Subcommittee on Competition Policy, Antitrust, and Consumer Protection in 2023. Francis, a former deputy director in the Federal Trade Commission’s (FTC) Bureau of Competition, also advocated antitrust reform and, not incidentally, increased funding for the federal antitrust-enforcement agencies. At the same time, like Hovenkamp, he was crystal clear that he did not “recommend enacting AICOA.” 

His detailed critique—more than 80 pages of his testimony focused on AICOA—covered ground that should be familiar to readers of Truth on the Market. For example, sections of his testimony carried subheadings identifying several problems with AICOA’s approach to self-preferencing: 

‘Self-Preferencing’ Includes Many Desirable Practices

Banning Self-Preferencing Would Inflict Consumer Harms

A Ban Would Deter Product Improvements

A Ban Would Deter Platforms from Protecting Consumers

A Ban Would Challenge Some Free-to-Use, Ad-Supported Services

A Ban Would Threaten Closed Ecosystems

Francis also argued that AICOA’s scope appears arbitrary. And he raised concerns about the bill’s limits on data use, its access and interoperability requirements, and its “no-conditioning rule.” On the latter, he noted that the “provision amounts to a per se rule against technological product tying of a kind that modern antitrust has long—and very wisely—left behind.” 

AICOA’s Vertical Leap of Faith

Part of what’s striking about the new AICOA is how many deeply flawed assumptions it retains from versions that failed to gain traction in earlier Congresses. Part of what’s also striking is how arbitrary some of its departures from the alpha and beta test versions appear to be. 

For example, the new AICOA’s general hostility to vertical integration, self-preferencing, and default settings is not new. Neither are its interoperability and data-portability requirements. Interoperability means designing a system so that it can interoperate–at the least, exchange data–with other systems; data portability means enabling users to move “their” data from one service or platform  to another. Both can be useful. But the decades-long push to encourage, and now require, interoperability in electronic medical systems is an object lesson in potentially confounding factors, even in a context where there are clear policy benefits to interoperability and there is public funding behind its development. Mandatory interoperability and data portability restrictions can prove costly, risky, or counterproductive, as can other data restrictions contemplated by AICOA. They do not even guarantee the effective flow of information, much less competitive benefits.   

Each of these issues could warrant a much longer critique. For more—much more—on self-preferencing, I’ll simply steer readers to ICLE’s Issue Spotlight on “The Case for Self-Preferencing,” with work from ICLE scholars, academic affiliates, and others.  

As a related matter, we’ve had quite a bit to say about vertical integration, but the basic point starts with the Supreme Court’s stepwise recognition—from Continental T.V. Inc. v. GTE Sylvania Inc. in 1977, which rejected per se liability for vertical nonprice restraints, through Leegin Creative Leather Products Inc. v. PSKS Inc. in 2007, which rejected per se liability for vertical price restraints—that vertical integration is not generally anticompetitive, even if it can prove anticompetitive under particular facts and circumstances. 

Directionally, at least, that trend in antitrust law has followed the economic literature. For example, James C. Cooper, Luke M. Froeb, Dan O’Brien, and Michael G. Vita reviewed the theoretical and empirical literature on vertical integration and vertical restraints and found “a paucity of support for the proposition that vertical restraints/vertical integration are likely to harm consumers.” They recognize that harm is possible—that is, the welfare effects of vertical integration are theoretically ambiguous—and argue that empirical evidence is therefore critical to sound competition policy. Their review of that empirical literature “suggests that vertical restraints are likely to be benign or welfare enhancing.” 

Similarly, Francine Lafontaine and Margaret Slade reviewed the literature on exclusive contracts and vertical restraints and found that, when firms adopt vertical restraints, they typically improve product quality and service, benefiting consumers as well as producers. 

The point is not that vertical restraints should be per se lawful. They remain subject to antitrust scrutiny under the rule of reason—the fact-specific legal test courts typically use to weigh competitive harms against benefits. Rather, as Cooper et al. observe, novel legal presumptions against vertical restraints are likely to produce too many false positives: condemning conduct that is actually benign or beneficial. Nothing in the literature, or in the “fact sheet” accompanying AICOA, suggests that the bill’s vertical restrictions will fare any better. 

What’s New? The Target—and the Hammer

So what’s at least somewhat new? Start with the target: Which firms would be subject to the AICOA’s restrictions?

The bill’s reach has been somewhat redrawn, although AICOA 1.2 retains the big-tech-is-bad approach of versions 1.0 and 1.1. It keeps the DMA-style “gatekeeper” model Hovenkamp criticized in assessing AICOA 1.0, along with the scope restrictions Francis found arbitrary in his testimony on AICOA 1.1. 

The new target is any “systemically important platform,” along with the firms that own or operate such platforms—or that hold, control, or benefit from at least a 25% share of them. It’s still Big Tech, but with a somewhat different metric for “big.” 

Under the bill, “big” turns partly on the scale of the parent or “operator”—the firm that owns or controls the online platform. AICOA targets firms with “average annual gross revenues of not less than $175,000,000,000,” as adjusted. It also turns on the platform’s popularity: “monthly active users in the United States equal to not less than 34 percent of the population of the United States over the age of 12, as determined by the most recent decennial census of population conducted by the Bureau of the Census.” Alternatively, the bill reaches platforms with monthly subscriptions held by the same percentage of U.S. households. 

To be sure, the revenue number is large. So is the monthly-active-user threshold. The most recent decennial census was conducted in 2020. It found roughly 280 million people over age 12. Multiply that by 0.34, and the threshold comes out to about 95.2 million monthly users. 

That is not just the GAFAM firms—Google, Apple, Facebook, Amazon, and Microsoft—and their platforms. Walmart and others would seem to fit the bill. Still, AICOA would reach a relatively small number of firms that operate online platforms. 

But why gross revenue? More broadly, why these large firms and these restrictions? Once again, there is no good answer—or even an intelligible one. Certainly, and consistent with earlier versions of AICOA, there is no market-power or monopoly-power requirement for any product or service market. 

Another new wrinkle is the bill’s civil-penalty provision: 

to be deposited in the Treasury of the United States, in an amount not greater than 10 percent, and not less than 1 percent, of the total United States revenue of the person for the period during which the violation occurred. 

That is 10% of revenue, not any measure of profits. And it is the revenue of the parent firm, not the “systemically important platform” itself. The 1% floor is still revenue, and it is still the parent firm’s revenue. As Manne says, that is “a new floor for enormous fines, even in close cases involving unsettled legal questions.” 

More than that, the bill contemplates civil penalties wholly untethered from any assessment of the harm supposedly caused by the prohibited conduct. Out the window goes the possibility of efficient remedies—that is, remedies that force firms to internalize the harms they cause and thereby give them incentives to avoid causing such harms in the future. 

Instead, we get runaway penalties that might be appropriate for unequivocally harmful conduct, where courts and enforcers have little concern about overdeterrence or penalties that ultimately do more harm than good to competition and consumers. That is not the world we are in here. None of this is about unequivocally harmful conduct. Compare and contrast, for example, Steven Shavell with Louis Kaplow and Robert Cooter. 

What of Manne’s twin complaints that AICOA “lowers the competition-harm threshold to anything above de minimis,” while increasing “the burden for key safety, privacy, and security defenses”? The first objection is lifted straight from the bill’s text. Key provisions—such as the self-preferencing provision—address conduct that “would materially harm competition.” And the definitions section sharpens the point: Under AICOA, “materially harms competition” means “any actual or reasonable risk of lessening competition or impairing the competitive process that is more than a de minimis amount.” 

Couple that very low bar with AICOA’s burdens on defendants seeking to establish an affirmative defense after the government pleads any risk of anything more than de minimis harm. For the prohibitions on “preferencing, limiting, and discrimination,” a defendant must show, by “clear and convincing evidence,” that the conduct was necessary either to comply with federal or state law or to “protect safety, user privacy, the security of nonpublic data or of the platform, or any other significant cybersecurity risk, or to prevent fraud or spam.” The defendant must also show that the conduct was “narrowly tailored in scope” and “could not be achieved through less anticompetitive means.” 

How are we to read “less anticompetitive means” against a standard triggered by any risk of anything more than a de minimis amount of harm? 

What’s more, any such affirmative defense would rely on the defendant’s production of ordinary-course documents that:

i. describe the specific purpose for which the conduct was undertaken; and

ii. identify the material risks or harms the conduct was intended to address.

For other prohibitions in the bill, the defendant would have to establish, “by a preponderance of the evidence that the conduct has not materially harmed and would not materially harm competition.” But again, material harm to competition means “any actual or reasonable risk of lessening competition or impairing the competitive process that is more than a de minimis amount.” 

These are not minor adjustments to the rule-of-reason scrutiny typically applied to allegedly anticompetitive restraints, and especially to vertical restraints alleged to violate Section 2 of the Sherman Act, at least since the Supreme Court’s 1977 decision in Continental T.V. Inc. v. GTE Sylvania Inc. 

Rather, for conduct likely to be procompetitive or benign, AICOA presumes illegality—setting the plaintiff’s burden very close to ground level. It then shifts the burden to the defendant, where the bill’s sponsors seem to have in mind “ain’t no mountain high enough”—a ’60s classic, to be sure, but not a theory of harm to competition or consumers. 

Third Time’s No Charm

The sponsors of AICOA 1.2 seem to have made at least a passing attempt to revise the details of versions 1.0 and 1.1. What drove those editorial choices is anyone’s guess. The key faults remain, and where the bill does change, it is hard to see a net improvement. 

This version flunks the “no economic sense” test, just like the last one, and just like the one before that. Herbert Hovenkamp said that “AICOA [1.0] was a bill that deserved to die.” That assessment was strict but fair. Entirely fair. 

So, with apologies to Sens. Klobuchar and Grassley: This one, too, deserves to die. Third time’s no charm.

The post AICOA Rises from the Grave, Still Looking for a Theory of Harm appeared first on Truth on the Market.

Last month, Brazil’s antitrust authority, the Administrative Council for Economic Defense (CADE), entered that debate. It issued part of its long-awaited decisions on a batch of AI partnership agreements and “acqui-hire” deals between Big Tech firms and AI startups. The practical result was modest. The doctrinal signal was not. 

CADE dismissed three cases: NVIDIA/Run, Microsoft/Mistral, and Google/Character.AI. It ordered ex post notification in one case, Microsoft/Inflection, and opened two new administrative proceedings involving Google/Windsurf and Google/Hume AI. One final case, Amazon/Anthropic, had been scheduled for decision at the same May Tribunal session, but Commissioner-Rapporteur José Levi Mello do Amaral Júnior withdrew it from the agenda, citing “notorious facts” that warranted further investigation. 

The headline result was prudent. Three cases escaped formal notification, and the one that did not—Microsoft/Inflection—had already been cleared on the merits by the United Kingdom’s Competition and Markets Authority (CMA). But beneath that restrained bottom line, the Tribunal opened two doctrinal doors that deserve careful scrutiny. 

The first is substantive. CADE held that “reverse acqui-hires”—bundles of nonexclusive licensing, key-team hiring, and substantial payments—can qualify as “concentration acts” under Article 90, II of Brazil’s Competition Law, even when no shares or direct assets change hands. In plain English, CADE signaled that a deal need not look like a conventional merger to be treated like one. 

The second is procedural. CADE invoked its “call-in” power under Article 88, §7º to require ex post notification of a below-threshold deal, even though the agency still lacks clear, predictable criteria for when it will exercise that exceptional authority. 

Both moves echo the July 2024 joint statement by the U.S. Justice Department (DOJ), Federal Trade Commission (FTC), European Commission, and the CMA on competition in generative-AI foundation models and AI products. CADE has, in effect, decided to surf the same wave. 

This post explains how it got there. It starts with the basics of Brazil’s merger-notification regime, then examines how the Tribunal characterized “acqui-hire” and “reverse acqui-hire” arrangements as “concentration acts.” It then asks what distinguished Microsoft/Inflection from Google/Character.AI on the “call-in” question, summarizes the outcome of each case, and closes with a substantive critique of the killer-acquisition theory CADE has imported. 

The question, in the end, is whether Brazil’s Competition Law has the doctrinal board to ride that wave safely. 

The Merger Review Before the Merger Review

CADE opened the five APAC proceedings—short for Apuração de Ato de Concentração, or investigation of a concentration act—in the second half of 2024, shortly after the joint statement from the four other leading enforcers. 

A brief procedural note helps. APACs are governed by CADE Resolution 24/2019. They are used mainly in gun-jumping cases—that is, cases involving transactions that should have been notified before closing but were not—and to determine, at a threshold level, whether a transaction qualifies as a “merger” under Brazil’s Competition Law. 

An APAC does not involve a substantive competition analysis of the transaction. At the end of an APAC, CADE’s Tribunal may order formal notification of the merger, which triggers the Ato de Concentração (AC) procedure. That is the proper vehicle for substantive legal and economic review. Under the AC procedure, CADE’s General Superintendence conducts the competition analysis and, depending on the case, refers it to the Tribunal for a final decision. 

That means none of these AI-related proceedings has yet received a meaningful substantive competition review. The reason is simple: an APAC is not built for that job. 

Instead, CADE’s Tribunal faced three preliminary questions. First, does each transaction qualify as a “merger” under Brazil’s Competition Law? Second, if so, does it meet Brazil’s mandatory notification thresholds? Third, if it does not meet those thresholds, should CADE nonetheless require formal notification under its “call-in” power?

Even without a full merger analysis, the APAC proceedings surfaced a long list of questions and hypothetical concerns worth examining. They also share one central legal fact: CADE opened them ex officio—on its own initiative—because none of the parties self-notified. And none self-notified because none of the deals met Brazil’s mandatory filing thresholds. 

The Thresholds Are the Point

Before turning to the specific cases, it is worth pausing over Brazil’s merger-notification system under Brazilian Competition Law (BCL) Law 12.529/2011. 

Under Article 88 of the BCL, a transaction must be notified to CADE before closing only if it first qualifies as a “concentration act” under Article 90. That category includes mergers, acquisitions of control or parts of companies—including tangible or intangible assets—and the contested category of “associative contracts,” which mainly covers joint ventures and similar arrangements.

Article 90’s concept of a “concentration act” is broad by design. Congress did not want merger control to be trapped by the narrower labels of business or contract law. Still, the category does not cover ordinary day-to-day commercial contracts between firms. A supply agreement is not a merger merely because the parties shook hands with gusto. 

Nor is classification as a “concentration act” enough to trigger CADE review. The transaction also must meet the revenue thresholds in Article 88. Mandatory notification arises only when both conditions are met: one party’s Brazilian group—typically the acquirer—recorded revenue of at least 750 million Brazilian reais (about $144 million) in the year before the transaction; and the other party’s group—typically the target—recorded revenue of at least 75 million Brazilian reais (about $14.4 million). If those thresholds are not met, the transaction is not subject to mandatory notification. 

The thresholds are jurisdictional. Without them—and absent a “call-in” decision—CADE has no authority to review the transaction at all. In that respect, Brazil is closer to the European Union Merger Regulation and its jurisdictional rules than to the U.S. Hart-Scott-Rodino Act regime, which is reporting-based rather than jurisdiction-conferring. 

Article 88, §7º creates the exception: CADE’s so-called “call-in” power. The statute provides that, “within one year from the date of consummation,” CADE may require notification of a transaction that otherwise falls below the thresholds. CADE has used that power sparingly. One study covering 2011—when the current BCL took effect—through 2023 found only seven uses. For perspective, CADE’s public dataset records at least 4,358 merger filings from 2015 through 2023 alone. Even that understates the full picture, because it excludes 2011 through 2014. 

Until now, the call-in power has been understood as a narrow corrective, reserved for deals whose economic significance slips through the revenue test. The five AI cases test just how far that correction can stretch before it becomes something else. 

In each of the four AI deals decided at the Tribunal’s May session, CADE’s General Superintendence issued its opinion in March of this year. It took nearly two years for the Superintendence to conclude that none of the transactions was subject to mandatory notification under Brazil’s merger-control regime. The acquirers—NVIDIA, Microsoft, and Google—comfortably cleared the 750 million Brazilian reais threshold. The targets—Run, Mistral, Character.AI, and Inflection—generated no meaningful revenue in Brazil. 

That resolved the second question identified above: The transactions did not meet the mandatory notification thresholds. The Tribunal then turned to the remaining two questions. First, do these transactions qualify as “mergers” under Article 90 of the BCL? Second, as the Article 88, §7º inquiry was reframed, should the targets’ lack of Brazilian turnover be overridden because AI startups may have strategic value?

Those are interesting questions. But especially on CADE’s call-in power, they raise a more basic one: Are these the right cases in which to answer them? And is this the best use of CADE’s resources when no agency has found evidence of actual harm to competition in these deals? 

The Amazon/Anthropic case adds another wrinkle. As noted above, CADE postponed that case for further investigation because of “notorious facts” that were not disclosed publicly during the Tribunal session. Why? Which facts or conditions justify different treatment? 

Other agencies have not found concrete competitive harm from the Amazon/Anthropic partnership. The FTC and the CMA both examined the deal. The CMA cleared it outright. The FTC’s report—conducted under the Biden administration and approved 5-0—identified potential risks, including lock-in, input foreclosure, and information asymmetry, but stopped short of enforcement action. 

When a Hiring Spree Starts Looking Like a Merger

In short, CADE’s Tribunal concluded that acqui-hires and reverse acqui-hires can qualify as mergers under Article 90 of the Brazilian Competition Law and may therefore require mandatory notification when Article 88’s revenue thresholds are met. That conclusion is significant. But the reasoning behind it—and the factual elements needed to turn a bundle of contracts into a “concentration act”—remains somewhat murky. 

As Selcukhan Ünekbas has explained, “acquihires are transactions aimed primarily at acquiring a firm’s workforce, rather than its products or other assets. They differ from ‘license-and-hire’ agreements, in which the acquirer also licenses the target’s technology.”

The controversy is not new. It traces at least to a paper by John Coyle and Gregg Polsky published more than a decade ago. So the debate over how Big Tech firms acquire, absorb, or otherwise siphon talent from startups—and what that means for competition law—has been around for a while, at least in the literature.

The newer label, or perhaps the newer lens, is the “reverse acqui-hire.” A traditional acqui-hire involves the direct acquisition of a small startup, with the workforce as the main prize. A reverse acqui-hire tries to reach much the same practical result without formally acquiring the company.

Reverse acqui-hires typically combine three sets of obligations, often spread across several contracts:

(1) some compensation paid to the team—conventionally structured to create incentives for the team to stay with the acquirer for a longer period of time; (2) a non-exclusive license to the startup’s intellectual property; and (3) some kind of waiver from the target company that it will not sue the acquirer for poaching the talent.

Unlike a traditional acqui-hire, the startup generally survives the deal: “In contrast to traditional acquisitions or acqui-hiring, RAH transactions neither involve equity transfer nor full buyout of technology nor full takeover of the people.” The startup usually reshapes its business afterward, but it remains a distinct entity, in some form, from its would-be acquirer. 

In Google/Character.AI, Commissioner Camila Alves examined the hallmark features of a reverse acqui-hire and how they bear on Article 90’s definition of a “concentration act.” In her view, the transaction should not be assessed as a pile of disconnected pieces, but as an integrated whole. Licensing intellectual property, hiring part of a technical team, terminating a prior investment, and making a substantial payment to the company might each fall outside Article 90 if viewed in isolation. But when bundled together as “part of a single economic package” (§ 59), CADE may treat the operation as a concentration act under Article 90, II of the BCL—either as the acquisition of “intangible assets” or as the acquisition of a “part of a company” (§ 72). 

Alves stressed that this is not a broad rule making every nonexclusive technology license or hiring decision notifiable. Rather, in her view, what matters is the coordinated structuring of those elements into a single negotiated package. Put less delicately: one contract may be a hire, another may be a license, and a third may be a payment. Together, CADE may see a merger wearing a fake mustache.

Commissioner José Levi reached a similar conclusion in Microsoft/Inflection. Microsoft argued that none of the deal’s individual elements—the hiring of Inflection co-founders Mustafa Suleyman and Karen Simonyan, the hiring of a substantial part of Inflection’s team, the licensing arrangements, and the related corporate contracts—amounted on its own to the acquisition of Inflection’s business or assets. José Levi agreed that the pieces should not be viewed in isolation. Taken together, he concluded, they fit Article 90, II.  

For José Levi, Microsoft/Inflection matched the reverse acqui-hire pattern described in the literature (§§ 46-51): leading technology firms effectively absorb a startup’s business activity by hiring its key personnel and compensating investors through substantial licensing payments, thereby producing the economic effect of a conventional acquisition outside the formal channels of corporate law. Because these arrangements replicate the economic logic of a conventional acquisition, CADE cannot decline jurisdiction merely because the transaction used an unconventional form. 

CADE therefore characterized Google/Character.AI and Microsoft/Inflection as reverse acqui-hire transactions based on the combination of elements in each deal, even though neither involved a direct acquisition of shares or assets. 

The other two cases differed. NVIDIA/Run was a straightforward full acquisition of Run by NVIDIA. Microsoft/Mistral involved a €15 million investment in instruments convertible into shares that, upon conversion, would represent 0.31% of Mistral’s equity. Under Articles 9 and 10 of CADE Resolution 33/2022, acquiring more than 5% of a company’s shares—combined with the Article 88 revenue thresholds discussed above—triggers mandatory notification. The acquisition of call options is treated as equivalent to a direct acquisition of shares for merger-control purposes. Those two cases therefore did not raise the acqui-hire questions that drove the analysis in Google/Character.AI and Microsoft/Inflection. 

To be sure, Article 90, II of the BCL is exceptionally broad, as noted above, and deliberately so. It provides: 

Art. 90. For the purposes of Article 88 of this Law, a concentration act shall be carried out when: […] II—one (1) or more companies acquire, directly or indirectly, by purchase or exchange of stocks, shares, bonds, or securities convertible into stocks or assets, whether tangible or intangible, by contract or by any other means or way, the control or parts of one or more companies.

Given that language, it is defensible to read reverse acqui-hire arrangements—combining personnel hiring, nonexclusive licensing, and monetary payments—as a single concentration act when those elements occur together, even though none would qualify on its own. 

The harder question is what, exactly, the contracts must contain before CADE can call the package an acqui-hire and, by extension, a concentration act. Is hiring one person enough, if that person is the chief executive officer? Must the acquirer hire part of the team, a substantial share of the team, or nearly everyone? What obligations must the nonexclusive licensing agreement include? Is an upfront payment necessary? If no compensation changes hands, can the transaction still be an acqui-hire? Must the contracts include a no-sue waiver by the target for talent poaching? And if one element is missing while the others are present, does the transaction still count? 

Those questions remain unresolved. That uncertainty is the predictable consequence of defining the concept by combining elements that, standing alone, would not trigger merger review. CADE’s Tribunal will need to refine these factors in future cases if companies are to have the legal certainty they need when deciding whether a transaction qualifies as a merger under Brazilian competition law. 

CADE Calls One In, Lets the Others Surf By

With the first question resolved (reverse acqui-hires can qualify as “concentration acts” under Article 90 of the Brazilian Competition Law) and the second settled as well (none of these AI partnership transactions triggered mandatory notification because they failed to meet Article 88’s revenue thresholds) the analysis turns to the third: CADE’s “call-in” power under Article 88, §7º. 

In the end, CADE required formal notification only in Microsoft/Inflection. The remaining cases were not called in. Why? 

One caveat: The analysis that follows focuses mainly on Microsoft/Inflection and Google/Character.AI, because the opinions in the other cases were not yet public at the time of writing. 

Article 88, §7º does not give CADE a roving license to second-guess every transaction—or every “concentration act”—that falls below the mandatory revenue thresholds. As both rapporteurs acknowledge, invoking the provision requires a juízo de conveniência e oportunidade: a discretionary assessment of whether a call-in is convenient and opportune in the specific case. 

Commissioner Alves frames the requirement clearly in Google/Character.AI: 

The provision (call-in power) functions as an exceptional mechanism for correcting the mandatory notification system. The revenue criteria confer objectivity, predictability, and administrative rationality on merger control, but they do not exhaust all hypotheses of possible competitive relevance. […] This possibility does not authorize the conversion of Article 88, §7º into an ordinary instrument for broad review of every non-notifiable operation. Its application requires caution, specific reasoning, a demonstration of competitive plausibility, and observance of the legal limits. It is an exceptional jurisdiction to be exercised when the circumstances of the case indicate that the ordinary revenue criteria may not have adequately captured the economic relevance of the operation. (§§ 21-22)

Alves reiterates the point in her conclusion: Article 88, §7º must be preserved “as an exceptional instrument for correcting the revenue-based notification system, and not as a mechanism of ordinary use in any atypical operation” (§ 133). On the standard itself, Alves (§§ 21-22 and § 133) and Commissioner José Levi (§ 60) appear to agree: The call-in power is exceptional; the transaction must present a plausible competitive concern; and the Tribunal must show that the ordinary thresholds failed to capture the deal’s economic significance. 

So what made Microsoft/Inflection convenient and opportune in a way that Google/Character.AI was not? Reading the two opinions side by side, three distinguishing features emerge—though the line they draw is hardly neon-bright. 

The first is the intensity of the team transfer. Alves flags the distinction: “the precedent (the Microsoft/Inflection case analyzed by the CMA in 2024) is not identical. In Inflection, the team transfer was more intense, involving almost the entire team of the company; in Character.AI, Google itself reports that [REDACTED] employees were released, in addition to a non-exclusive license over certain technologies” (§ 68). 

José Levi makes a similar point in Microsoft/Inflection, relying on the CMA’s parallel findings: “from the data in the CMA’s decision on the same operation, it is possible to infer that the employees hired by Microsoft corresponded, in fact, to almost the entire team of Inflection” (§ 72). The redacted portion of Camila’s analysis suggests Google hired a meaningful subset—but not the whole team—of Character.AI. The scale of the team transfer therefore appears to have mattered to CADE’s decision to require formal notification in Microsoft/Inflection. 

The second distinction is what happened to the target’s product strategy after the deal. This is the point on which José Levi rests most heavily. Before the transaction, he explains, Inflection focused on developing foundation models—general-purpose AI models that can support many downstream applications—and used them as the basis for Pi, its consumer-facing chatbot. After the team responsible for growth in that segment departed, Inflection allegedly shifted its focus to customized generative AI models for corporate clients (§ 75). 

José Levi treats that pivot from business-to-consumer to business-to-business as “circumstantial evidence that the operation could have produced a reduction of rivalry in the segment of FM development and consumer-facing chatbots” (§ 76)—a segment where Microsoft operates through Copilot. Character.AI, by contrast, continued operating its consumer-facing chatbot platform after the Google deal. As Alves notes, it “preserved formal independence” and “continued its downstream application” (§ 140). 

The third distinction is horizontal overlap. José Levi’s emphasis on Microsoft Copilot and the consumer-facing foundation-models segment frames Microsoft/Inflection as a deal that may have reduced current rivalry, not merely future rivalry, in a market where the acquirer already competes (§ 77). Alves’ analysis of Google/Character.AI leans more heavily on potential competition and innovation trajectory—the idea that Character.AI might have become a more significant rival in the future (§§ 81-82, 138-139). Whether the Tribunal believes this distinction matters under Brazilian Competition Law remains unclear. 

Despite these differences, Alves is explicit that the dismissal of Google/Character.AI rests not only on the absence of competitive plausibility, but also on timing and proportionality. She writes:

I understand that it is neither convenient nor opportune to determine the notification of the operation. This conclusion does not derive exclusively from the time elapsed since consummation, although that element is relevant in the exceptional design of Article 88, §7º. It also weighs the convenience of an ex post requirement at this moment, the already advanced stage of implementation of the operation, the absence of consolidated administrative criteria for reverse acqui-hires, the proportionality of the measure, and the possibility of directing institutional action toward more recent analogous operations, in a specific and timely procedure. (§ 141).  

José Levi acknowledges similar factors in Microsoft/Inflection but reaches the opposite conclusion. He leans instead on the killer-acquisition and talent-hoarding literature (§§ 65-69), the intensity of the team transfer, and the alleged reduction of rivalry in the consumer-facing foundation-models segment. 

A skeptical reader is entitled to find the resulting line less than crystalline. Both deals were announced and consummated in 2024. Both involved nonexclusive licenses, substantial payments, and the hiring of part—or nearly all—of a target’s technical team. Both were treated as “concentration acts” under Article 90, II. And under Alves’ opinion, both were at least plausibly relevant from a competition standpoint. 

The different outcomes therefore seem to rest less on a clear legal rule than on a bundle of factual gradations and ex post implications: how many team members were hired, what happened next at the target, and how closely the target’s business overlapped with the acquirer’s existing products. 

That may be enough to decide these cases. It is less helpful for the next company trying to plan around CADE’s call-in power. 

Based on the decisions so far, it remains difficult to articulate a clear legal test for when CADE will require formal notification. For companies assessing whether their own license-and-hire arrangements may attract scrutiny, the emerging criteria remain opaque. CADE’s Tribunal will need to clarify them in future cases—and it will soon have the chance to do so in Google/Windsurf and Google/Hume AI. 

CADE Catches the AI Wave—Now Comes the Hard Part

After this brief tour of the cases, the outcomes were as follows. 

In NVIDIA/Run, Commissioner-Rapporteur Carlos Jacques Vieira concluded that Run does not operate in Brazil, that the transaction does not produce effects in Brazil, that the parties lack market power in the country, and that there was no harm—or risk of harm—to competition sufficient to justify invoking Article 88, §7º. The European Commission reviewed the same case in December 2024 and approved it, finding that it raised no serious doubts and would not harm consumers. Like the other cases discussed below, NVIDIA/Run was decided unanimously by the four current members of CADE’s Tribunal. 

In Microsoft/Mistral, Commissioner-Rapporteur José Levi found that Microsoft’s stake in Mistral fell below the 5% threshold, that the partnership agreement did not confer control, and that there was no evidence of competitive harm in the relevant market. He therefore dismissed the case and declined to require formal notification under Article 88, §7º. That outcome tracks the analysis and predictions advanced by Dirk Auer and Mario Zúñiga when the investigations were first opened. The CMA had reviewed the same transaction in May 2024 and similarly concluded that, given the deal’s specific features, “the CMA therefore does not have jurisdiction to review the Partnership in its current form.” 

In Google/Character.AI, Commissioner-Rapporteur Alves voted to dismiss but wrote a substantial opinion on AI partnerships and acqui-hiring agreements. As discussed above, Alves treated the combined deal—a nonexclusive technology license, the “release” of part of Character.AI’s technical team for Google to hire, the cancellation of a prior investment, and a payment of roughly $2.7 billion—as a “concentration act” under Article 90, II of the BCL. She nonetheless declined to require formal notification after weighing the costs and benefits of an ex post filing requirement. 

In the same decision, Alves ordered CADE’s General Superintendence to open two new APAC procedures involving the more recent Google/Windsurf and Google/Hume AI cases. CADE’s debate over whether acqui-hiring arrangements may qualify as concentration acts under Brazilian merger-control law is therefore very much alive. 

Finally, in Microsoft/Inflection, José Levi required formal notification. As discussed above, he treated the transaction as following the same basic template as Google/Character.AI: the hiring of nearly the startup’s entire team, combined with a nonexclusive license and a substantial upfront payment. He found that package to be a “concentration act” under Article 90, II of the BCL. Microsoft must now formally notify the transaction under Article 88, §7º within 30 days. 

CADE will therefore have to assess the transaction’s substantive effects, including the relevant market, the degree of market concentration, and the potential for unilateral and coordinated effects. It will also likely have to confront, at least to some extent, the “killer acquisition” theory of harm—the claim that an incumbent buys, absorbs, or neutralizes a nascent rival to prevent future competition. 

That is notable because CADE required formal notification even though the CMA had already cleared Microsoft/Inflection on the merits in September 2024, finding no realistic prospect of a substantial lessening of competition. Microsoft/Inflection will thus become the first AI partnership agreement to receive full review under CADE’s regular Ato de Concentração merger-control procedure. 

CADE’s first wave of AI partnership decisions produced a restrained practical outcome but an ambitious doctrinal framework. The Tribunal expanded the interpretation of concentration acts to encompass reverse acqui-hires and signaled a greater willingness to use its exceptional call-in power for transactions involving AI startups. 

Those uncertainties are not merely procedural. They reflect a deeper substantive debate now emerging internationally and increasingly shaping Brazilian competition law. The next question is whether the theories driving this enforcement agenda—especially the theory of digital killer acquisitions—rest on solid empirical ground. 

That question is the focus of the second article in this series. For now, CADE has paddled into the AI wave. The harder part is proving it can steer.

The post Brazil Catches the Acqui-Hire Wave appeared first on Truth on the Market.

To many observers, this looks wasteful at best and sinister at worst. Why would a profit-maximizing firm pay enormous sums for talent it seemingly has no intention of using?

Ronald Coase had a wry answer for moments like this. “[I]f an economist finds something—a business practice of one sort or other—that he does not understand,” he observed in 1972, “he looks for a monopoly explanation. And as in this field we are very ignorant, the number of ununderstandable practices tends to be rather large, and the reliance on a monopoly explanation, frequent.”

A new working paper by Shaolong Wu of Harvard Business School and Zefan Qian of Georgetown, “Talent Hoarding and Upstream Innovation: Labor Market Distortions by Large Incumbents,” supplies precisely that monopoly explanation, complete with a formal model and an empirical test. Large incumbents, the authors argue, sometimes hire and retain frontier researchers not to put them to work, but to keep rivals from doing so. Because top AI-research talent is scarce, every researcher a dominant firm keeps “on the bench” is one a challenger can’t hire. The incumbent protects its existing profits, and society loses the discoveries that researchers would have produced elsewhere.

The paper closes with policy recommendations to match: limits on garden leave, narrower noncompete agreements for publicly funded researchers, and institutional pressure to keep frontier talent “actively deployed.”

The policy audience is already primed for this argument. The Federal Trade Commission (FTC) has announced its intention to scrutinize acquihires—transactions structured around hiring a startup’s employees rather than acquiring the company outright—to ensure they aren’t used to evade merger review. Talent-centered theories of competitive harm are rapidly becoming the next front in the broader campaign against large technology firms. All the more reason to get the economics right.

It’s a clever paper, and perhaps a more careful one than many in this genre. Its headline claim, however, substantially outpaces its evidence. What the data actually show is that one group of software firms retained more skilled employees after the Supreme Court weakened their patent protections. Everything beyond that—the “idle benches,” the foreclosed rivals, the lost innovation, and the social harm—comes not from the data but from assumptions built into the model.

More importantly, nearly every one of those assumptions rules out, by construction, a far more ordinary explanation: that the same behavior reflects good management rather than anticompetitive conduct.

The Model Does the Hoarding for You

In the paper’s model, an incumbent firm earns profits from a legacy technology, while an entrant may develop a replacement. The key input into discovering that new technology is a fixed pool of researchers (an assumption we will return to) who possess knowledge of the incumbent’s technology. Hiring one of those researchers does double duty: It increases your own odds of a breakthrough while reducing your rival’s. There is, by assumption, no one else to hire. 

The model also imposes another crucial condition. A researcher generates innovation only when paired one-for-one with “innovation capital”—new equipment, computing power, organizational capacity, and similar inputs. A retained researcher who lacks that matching investment is “benched”: employed and paid, but contributing nothing to the only outcome the model treats as socially valuable—the discovery itself. The model does allow benched workers to perform ordinary legacy work—it simply assigns that work no social value.

An additional assumption drives the paper’s headline results. Those results emerge from what the authors call the “no self-replacement” case, in which a breakthrough is assumed to be worth no more to the incumbent than the legacy product it already sells. The incumbent, therefore, gains nothing from innovation itself. In such a world, where the breakthrough has no upside for the incumbent, the only reason to retain a researcher is to keep her away from a rival. 

Notably, retention doesn’t require leaving the researcher idle. Matching her with capital and putting her to work also reduces the entrant’s chances of success. Under the model’s assumptions, even productive employment can function as a form of exclusion. Whether she sits on a bench or works at a desk, the point is the same: she isn’t helping a competitor build the next generation of technology.

The paper thus assumes, quite literally, that an incumbent’s only reason for incurring the cost of retaining a knowledgeable researcher is to deny that researcher to a rival. Under those conditions, if a successful entrant would destroy the incumbent’s profits, the incumbent will rationally pay to keep researchers in-house even when it places no value on anything they might produce.

On the empirical side, the paper exploits the U.S. Supreme Court’s 2014 decision in Alice Corp. v. CLS Bank. The Court held that implementing an abstract idea on a generic computer is not, without more, patentable. The decision thus made broad software patents substantially harder to obtain and enforce. For firms whose competitive position depended on such patents, Alice increased the risk that rivals could imitate or displace them—what the paper terms “external displacement risk.” Talent hoarding, in turn, became a potentially more attractive way to protect an incumbent position.

The authors find that software firms holding pre-2014 software patents increased their net hiring after Alice by roughly 0.34 percentage points per month relative to software firms without patents—a not-insubstantial effect, about 63% of the sample average. 

Several additional findings are noteworthy:

• The increase came almost entirely from reduced employee departures rather than increased recruiting.
• The effect was concentrated among workers with advanced degrees.
• It was not accompanied by measurable growth in assets or other complementary capital.
• A companion analysis of worker résumés found that science-and-research employees at exposed firms advanced more slowly up the occupational ladder in the years following the decision.

In short, the study finds: more high-skilled workers retained, no corresponding increase in complementary capital, and slower career progression among the researchers who stayed. The authors interpret this pattern as evidence of their central hypothesis: the defensive retention of underutilized talent.

The Data Are Not the Dispute

To give the paper its due, the distinction between retaining a worker and deploying one is genuinely useful. Too much commentary on tech hiring blurs the difference. The research design is also serious: exposure is measured before the shock, the shock itself is a judicial decision no firm controlled, and the authors are unusually candid about their limitations. Most importantly, the central empirical finding appears real. Patent-exposed software firms reduced employee separations after Alice, and those reductions were concentrated among workers with knowledge likely to be valuable for innovation.

The dispute isn’t over the finding, but over its interpretation. The talent-hoarding story depends on a series of modeling choices, each of which rules out a competing explanation under which the same data reflect efficient, rather than anticompetitive, behavior.

Foreclosure by Construction

The paper’s most significant assumption is one most readers will never notice. It has two parts. First, the incumbent and the entrant are assigned identical discovery technology, meaning a researcher generates exactly the same probability of a breakthrough wherever she works. The incumbent firm itself contributes nothing distinctive, except already being there. Second, as noted above, in the headline case, the incumbent places no value on producing the breakthrough at all. 

Taken together, the model effectively assumes the incumbent is the worst possible home for research talent. It’s no better at research, and it’s uninterested in the results. The only remaining explanation for retention is blocking. 

The authors would fairly answer that this isn’t a claim about the world. They set the new technology’s value to zero on purpose, to strip out the incumbent’s offensive motive—wanting the breakthrough for itself—and study the defensive motive in isolation. Fair enough. Isolating one mechanism is what models are for.

But here, the model removes the efficient reason to retain talent before the welfare verdict is rendered. Then that verdict gets carried over to a world where incumbents usually do want the breakthrough, and where retention is often offensive as well as defensive. The conclusion isn’t “found” in the research; it’s built into the model. 

And that assumption elides a great deal. Firms aren’t equally good at converting researchers into discoveries. They differ enormously in management quality, in the data and infrastructure researchers can use, in their ability to select promising projects and kill bad ones, and in the colleagues a new hire works alongside. A large incumbent may get more out of a marginal researcher than a startup can, because she joins teams with experienced managers who have already solved a thousand deployment problems, and because she can work with proprietary data and installed systems no entrant possesses. 

In that model—the real world—scarce talent flowing toward incumbents isn’t foreclosure. It’s the market allocating a scarce input to its highest-valued use. Eye-popping compensation is the competitive price of that input, captured by the workers themselves.

The paper’s model rules out this more realistic dynamic, and the empirical work can’t restore it. A finding that exposed firms retained more researchers is equally consistent with “those firms denied rivals an input” and “those firms are where the input is worth the most.”

An economist might respond that, in the model, the incumbent’s willingness to pay exceeds the worker’s deployment value. That’s what makes it hoarding. While true in this context, that result follows from the equal-capability assumption. Relax that unrealistic assumption, and the wage premium may reflect a productivity difference rather than a blocking premium.

That means the paper’s welfare conclusion turns on a parameter it never measures.

The Salop Problem Returns

The claim that a dominant firm will rationally outspend any challenger to preserve its position is one of the oldest moves in the antitrust playbook. A recent version of it comes from Steven Salop, who argued—in a paper pointedly titled “Potential Competition and Antitrust Analysis: Monopoly Profits Exceed Duopoly Profits”—that acquisitions of potential or nascent competitors by dominant firms raise inherent anticompetitive concerns. Because keeping a monopoly is worth more than sharing a duopoly, the incumbent will always pay more to eliminate a threat than the threat is worth to anyone else. 

Wu and Qian have, in effect, transported this logic from the market for startups to the market for researchers. Here, the thing the incumbent supposedly overpays to control isn’t a nascent rival firm, but the researcher who might help create one. Same wine, different bottle.

The argument fails in the labor-market setting for many of the same reasons it fails in the acquisition context—reasons Dirk Auer, Brian Albrecht, Eric Fruits, Daniel Gilman, Lazar Radic, and I discussed in the International Center for Law & Economics’ (ICLE) comments on the FTC and U.S. Justice Department’s (DOJ) draft merger guidelines, and that Albrecht has explored at length here at Truth on the Market.

The first problem is arithmetic. Buying off a single potential entrant may be profitable when monopoly profits exceed duopoly profits. But once one challenger is paid to stand down, the next stands to enter as a duopolist, and the next after that, and so on. Each must be compensated at roughly the duopoly level, not some small fraction of it. With enough potential challengers, the cost of paying them all off exceeds the value of preserving the monopoly, and the strategy collapses. If we nonetheless observe the conduct, that suggests something other than monopoly maintenance may be at work.

The same arithmetic applies to talent. The paper’s model contains one incumbent, one entrant, and one closed pool of researchers, so cornering the input requires outbidding exactly one rival. The real frontier-AI labor market looks nothing like that. There are multiple deep-pocketed AI labs, a venture-capital ecosystem aggressively funding startups to compete for the same people, and a university pipeline continuously producing new talent.

To foreclose discovery through retention, an incumbent would have to outbid all of those rivals, for every pivotal researcher, indefinitely. The predictable result isn’t foreclosure, but an auction. And in that auction, the owners of the scarce input—the researchers themselves—capture much of the surplus. That looks less like market failure than competition doing exactly what it is supposed to do.

Two features of labor markets make the strategy even less durable than its acquisition counterpart. First, workers, unlike startups, can’t be bought outright. Because “human capital is inalienable,” a retained researcher stays only as long as she chooses to stay. Maintaining exclusivity is therefore perpetually expensive and contractually fragile. 

Second, when employees leave large firms, they disperse. They don’t march single-file to the one rival most capable of threatening their former employer. As Kevin Murphy has observed, departing workers spread across many employers, with only a small fraction joining any particular competitor and many leaving the industry altogether. The model’s premise that every researcher released by the incumbent flows directly to the one firm poised to destroy it bears little resemblance to how labor markets actually function.

The second problem is the one discussed above. Salop’s result requires the incumbent to be at least as capable as any challenger—an assumption he relegates to a footnote, while acknowledging that “monopoly profits are not always higher” when an entrant has lower costs or a better product.

As Dirk Auer, Sam Bowman, and I have written:

Although it is convenient in theoretical modeling to assume that similarly situated firms have equivalent capacities to realize profits, in reality firms vary greatly in their capabilities, and their investment and other business decisions are dependent on the firm’s managers’ expectations about their idiosyncratic abilities to recognize profit opportunities and take advantage of them—in short, they rest on the firm managers’ ability to be entrepreneurial.

Once that realistic possibility is admitted, differences in firms’ willingness to pay no longer demonstrate preemption. They may simply reflect differences in productivity.

Again, the empirical prediction that retention rises after Alice doesn’t depend on the equal-capability assumption. But the conclusion that retention is driven by exclusionary motives does. The empirical test therefore confirms the portion of the model that doesn’t require the assumption, while leaving untested the assumption that does most of the work in generating the paper’s policy recommendations.

No New Assets, No New Ideas?

The paper’s cleverest empirical move is to infer foreclosure from the combination of rising headcount and flat capital investment. At first glance, the logic seems straightforward: if firms were truly putting these researchers to productive use, we should see corresponding investment in the inputs they need.

But the argument rests on a strong assumption: that a researcher paired with anything less than a full new unit of “innovation capital” produces nothing. In software, of all industries, that is a peculiar picture of the production process.

The complementary inputs that matter—data, codebases, proprietary tools, accumulated organizational knowledge, and much of the underlying computing infrastructure—are often already in place. An additional researcher can use those resources without the firm having to book a single new asset. Much of this capital is also non-rival within the firm. Another engineer working with the company’s data doesn’t consume that data. Even rivalrous inputs, such as computing power, are often available within existing capacity, with enough slack to absorb another researcher without generating a measurable increase in asset purchases.

A firm whose competitive advantage consists precisely of these intangible assets can productively absorb additional researchers with no observable capital response at all. Put differently, the paper’s headline empirical finding—more labor, flat balance-sheet assets—is exactly what we would expect if incumbents are where marginal researchers are most productive. 

The authors also argue that complementary capital “can often be scaled within just days and weeks,” making its absence particularly informative. If firms wanted to deploy these researchers, they argue, they could have done so cheaply and quickly.

But that cuts both ways. If complementary capital is as inexpensive and scalable as the authors suggest, then the absence of a detectable increase in assets is weak evidence of non-deployment. Cheap capital may not show up as a measurable balance-sheet change even when researchers are fully deployed. And if computing capacity can be expanded on short notice at low cost, then computing power was never the binding constraint. Talent was.

In that case, scarce talent flowing toward the firms willing to pay the most for it is not evidence of benching. It is what efficient allocation looks like.

When the same evidence supports opposite welfare conclusions, it ceases to be especially informative.

Nor does the authors’ fallback diagnostic—that patent output per employee fell at exposed firms—break the tie. That claim is perilously close to circular. The Supreme Court had just made it more difficult for those firms to obtain patents. Of course patenting declined. That was the treatment. It is not evidence that researchers were working less effectively or producing less valuable output.

The People Are the Patents Now

The most natural reading of the paper’s central finding requires no “benches” or anticompetitive animus. Alice didn’t conjure displacement risk from nowhere; it weakened the specific legal instrument—patents—that exposed firms had used to protect their innovations. Economists have long understood patents and trade secrecy as substitute ways to appropriate the returns to invention. And trade secrets don’t live in filing cabinets—they live in employees’ heads.

When the Court devalued these firms’ patents, the rational response was to lean harder on the other instrument: retaining the people who embody proprietary knowledge.

That reading fits the paper’s evidence rather neatly. It predicts retention rather than recruitment, because the point is to protect knowledge the firm already has, not acquire more of it. It predicts concentration among advanced-degree workers, because they are most likely to carry the relevant know-how. And it predicts no corresponding capital expansion, because nothing about the firm’s investment program or resource allocation needs to change. 

The paper’s own marquee example points in exactly this direction. Describing Adobe—the lead illustration—the authors write that, once Alice weakened its patents, “the strategic asset that mattered more was not just the code or the patent portfolio, but the people who embodied the know-how.” That’s a description of a firm switching from one appropriability mechanism (patents) to another (retaining employees who hold trade secrets). The paper states the benign reading in its own words, in its own flagship example, and then subsumes it under “foreclosure.”

But those are not the same thing. Preventing your own know-how from walking out the door to a competitor is a legitimate interest that trade-secret law has protected for well over a century. On the most plausible reading, it’s also welfare-enhancing here: retention substitutes for the appropriability that Alice weakened, preserving at least some of the innovation incentives the decision would otherwise have eroded. 

Indeed, the paper’s own framing concedes the point without quite noticing it. The workers in the model are valuable to the entrant because of their “incumbent-specific knowledge.” What the rival wants from them, in other words, is largely the incumbent’s own proprietary information.

The Missing Foreclosed Rival

Even granting the retention findings, the paper’s welfare conclusion—that society loses when incumbents retain these workers—is built into the model rather than derived from the evidence. The social planner against whom the incumbent is judged is defined to value “the availability of the innovation” while placing no weight on the incumbent’s existing profits. By assumption, entrant innovation is socially valuable. 

But a challenger’s incentive to enter often includes what economists call “business stealing.” Much of the entrant’s prospective profit comes from taking customers and profits from the incumbent. From a social perspective, that is largely a transfer rather than a gain. Indeed, the economic literature on entry has long recognized that private incentives to enter can sometimes exceed the socially optimal level for precisely this reason. 

One could just as easily write down a model with the opposite welfare weights and conclude that the entrant’s poaching is the problem. Neither stipulation would constitute evidence.

Meanwhile, the foreclosure side of the story—the part that carries the antitrust implications—is never tested at all. Foreclosure requires, well, foreclosure: rivals must be meaningfully constrained in their ability to compete because they can’t access a necessary input. Yet the paper offers no evidence that startups in Alice-affected fields were starved of talent, hired fewer workers, grew more slowly, or innovated less than they otherwise would have. Perhaps they did, but the paper neither tests nor demonstrates it.

What evidence we do have points in the opposite direction. The authors themselves note that the cost of software experimentation was falling during this period, as cloud computing and related technologies made entry cheaper and more accessible. 

Readers familiar with the killer-acquisitions debate will recognize the problem. Even in pharmaceuticals—the sector with the strongest evidence for the theory—the most generous estimates suggest that genuine “killer” acquisitions account for only about 5% to 7% of deals. The evidence in digital markets is considerably weaker—which is to say, nonexistent.

A foreclosure mechanism that leaves no observable trace in the supposedly foreclosed market may be an interesting theoretical possibility. It is not an empirical finding.

A Raise Is Not a Market Failure

The paper’s career-trajectory evidence is also clever, but again curiously interpreted. The slowdown in advancement for workers in the “science and research” category at exposed firms is real but tiny—about 0.018 points on a 2-to-5 occupational scale. Indeed, the scientists who are the paper’s central concern show the smallest of the highlighted effects. 

The largest decline in the paper’s table of “strong negative effects” belongs instead to workers in “education and training,” at a rate of ?0.068—nearly four times the science-and-research estimate. That is difficult to square with a story about the “benching” of frontier talent. 

The next-largest slowdown, among workers in “law, compliance, and public safety,” is easier to explain. The authors themselves attribute it to a collapse in demand for patent-related legal work—a direct consequence of weaker patents, not talent hoarding. And in a separate “difficult to interpret” category, “healthcare and maintenance” workers at exposed firms also show larger slowdowns than science-and-research employees. 

The heterogeneous responses across job categories are supposed to support the hoarding story. But doing so requires emphasizing the category with the smallest effect while explaining away the larger effects the theory can’t account for.

There is also a simpler, and perhaps more important, problem: the data contain job titles but not wages. A worker who stays in the same role because her employer pays her substantially more hasn’t suffered a policy-relevant harm. She has been compensated. Indeed, the very study the authors cite on inventors moving to large firms finds that earnings rise by double digits after those moves.

A flatter title progression accompanied by a fatter paycheck isn’t a labor-market distortion. It’s known as a raise.

Before We Ban Good Management 

If we step back and ask what conduct this paper actually condemns, the normative implications become difficult to sustain. What the paper identifies is lower employee turnover, the payment of retention premiums, and efforts to keep proprietary knowledge from reaching competitors. That describes competent management far more readily than monopolization. Every well-run firm in every industry does these things. The conduct labeled anticompetitive is observationally indistinguishable from conduct we generally want firms to undertake. And the paper’s empirical design can’t tell the difference, because the assumptions that do the distinguishing work—equal firm capability, one-for-one capital matching, and a social planner indifferent to incumbent investment incentives—are imposed rather than tested.

The paper’s reception has outpaced its findings. One economist described it as “a dark theory (with evidence)” of Big Tech hiring, in which incumbents “leave some on the bench” to “slow rival innovation.” But the bench is precisely what the paper never observes. The authors are candid that they can’t see deployment worker by worker. Benching is inferred from the absence of a capital increase, not directly measured. “With evidence” is doing a great deal of work for a mechanism the study assumes rather than demonstrates.

Despite that, the paper advances policy recommendations: limits on garden leave, restrictions on noncompete agreements for publicly funded researchers, and institutional expectations that frontier talent remain “actively deployed.” These proposals arrive amid a broader labor-antitrust push whose evidentiary foundations are already remarkably thin, resting on a small number of studies with heavily qualified findings. 

Yet the mechanisms the paper would restrict perform real economic functions. Even the FTC, in the rulemaking that banned noncompetes, acknowledged evidence that such agreements can increase investment in workers’ human capital, physical capital, and research and development. Firms invest in training and entrust employees with valuable knowledge when they have some confidence that those investments won’t immediately walk across the street. 

Weakening those arrangements based on a possibility theorem and an empirical pattern that admits several efficient explanations is a recipe for getting error costs backward. It risks condemning ordinary retention practices in the hope of catching the occasional genuine “idle bench.”

It is also worth remembering which firms the talent-hoarding story is meant to indict. The paper takes its empirical cue from a 2014 patent decision, but its stakes—and its policy recommendations—are aimed squarely at today’s AI incumbents. Those firms are not behaving like hoarders stashing talent on idle benches. From 2019 to 2025, the largest tech firms—Amazon, Alphabet, Meta, Microsoft, and Apple—added nearly 1 million employees while increasing capital expenditures from $77 billion to $370 billion, according to their 10-K filings. That is not a picture of firms accumulating talent they have no intention of equipping. It is a picture of firms racing to deploy labor and capital as quickly as they can acquire them. In other words, it looks like competition, not foreclosure. 

The paper’s actual contribution is narrower, but still interesting—if perhaps not as headline-grabbing. It suggests that when courts weaken patent protection, firms respond by retaining the people who embody their proprietary knowledge. That’s a finding about the substitutability of appropriability mechanisms, and it counsels caution both about Alice-style doctrinal shocks and about labor-market interventions that would weaken the substitutes firms turn to in response.

Before talent retention becomes the next frontier of antitrust, we should demand evidence of the thing that matters: not that firms kept their researchers, but that those researchers sat idle, and that someone else would have put them to better use.

The post Gotta Catch ’Em All? Antitrust and the AI Talent Wars appeared first on Truth on the Market.

On April 28, 2026, the European Commission published the review required under Article 53 and declared the DMA “fit for purpose.” The report credits the law with “a tangible positive impact” and sees no need to revise the list of core platform services. In the Commission’s telling, the machinery is working; where the evidence remains thin, time will supposedly fill the gaps. 

I do not doubt the Commission’s sincerity. I doubt that the exercise could ever have produced a different result. 

The DMA’s evaluative architecture is designed in a way that makes failure effectively impossible to demonstrate. Not because the Commission is uniquely stubborn or populated by self-serving sycophants. Rather, the regime’s foundational premises determine what counts as evidence in the first place. In a  new white paper, I describe this phenomenon as autopoietic regulation: a system that reproduces its own assumptions rather than testing them. The DMA’s first review confirms that diagnosis precisely because it passes. 

The Review That Couldn’t Fail

Start with a simple question: What finding would have led the Commission to conclude that the DMA, or some core part of it, was a mistake? 

Not “needs more time.” Not “compliance is incomplete.” A genuinely negative verdict: that an intervention’s costs exceeded its benefits; that a prohibition targeted conduct that was making consumers better off; or that the regulation’s underlying premise was wrong. Any review worthy of the name should be able to answer that question. An evaluation that can return only “working” or “working, but not yet finished” is not really an evaluation at all. 

The DMA’s review cannot return such a verdict, and I argue that the reason is structural. The regulation’s recitals present its obligations as simultaneously good for end users, business users, and innovation. In the DMA’s self-conception, there are no tradeoffs (except gatekeepers’ losses, which, as I have argued, are the point). 

No recital seriously entertains the possibility that mandated interoperability could reduce quality, that a prescribed fee structure could distort investment incentives, or that redistributing rents from platforms to business users could leave consumers—or other business users—worse off. If the statute does not acknowledge such costs, the review designed to assess the statute has no category in which to record them. 

So watch what happens when those costs show up anyway. 

When Bad News Doesn’t Count

The review did not lack for adverse evidence. The accompanying staff working document notes that advertisers, despite gaining new data-access rights, still face “persistent issues in obtaining sufficiently granular and comparable data.” It reports that alternative app stores and payment methods have produced “mixed results,” hampered by technical and contractual barriers. Gatekeepers and others also warned that interoperability mandates carry high compliance costs and “can harm customers, stifle innovation, and raise security and privacy concerns.” 

In an ordinary oversight regime, observations like these would mark the beginning of an inquiry. Has the intervention made its intended beneficiaries better off? Are the benefits worth the costs? Under the DMA, they function largely as consultation inputs: recorded, acknowledged, and then set aside. 

The recurring move is to treat disappointing results not as evidence that an obligation may be misconceived, but as evidence that it has not yet had time to work. App-distribution rules underperforming? Too soon to judge. Messaging interoperability barely adopted? Too soon to draw conclusions. The asymmetry is striking: success counts immediately, while failure is always provisional. 

The clearest illustration comes from the one company that does not fit the DMA’s template. Amazon—the only retailer designated under a framework built largely for search engines, app stores, and operating systems—used the review’s publication to argue publicly that the assessment “fails to address the fundamental design flaws that make the DMA unfit for retail,” and that compliance has “stalled investment, damaged the customer experience, and not made services more affordable.” 

One need not take Amazon’s word for any of this. The point is what the review does with such claims. It records them as stakeholder grievances. The possibility that a 40%-50% decline in a retailer’s organic traffic might reflect a defect in the regulation, rather than transitional friction on the way to a better equilibrium, lies outside the frame the Commission brought to the exercise. 

A Machine That Eats the Evidence

Here is where the autopoiesis point—big word, I know—connects to the architecture of the Digital Markets Act (DMA), and to two ideas I develop at length in the paper: the structural-unfairness axiom and the conflation of public and private gain. 

The DMA does not ask whether a gatekeeper’s conduct harms competition to the detriment of consumers. It begins from the axiom that designated markets are structurally unfair—that gatekeepers, by virtue of their position, “obtain a disproportionate advantage.” Unfairness is not something the Commission must demonstrate case by case. It is a premise the regime carries in. 

That kind of premise is self-sealing. Whenever an asymmetry between gatekeepers and business users persists—whenever rivals have not gained share, or users have not migrated—the axiom supplies a ready explanation: The unfairness has not yet been corrected. There is no result the premise cannot absorb, because the premise was never stated in a form that data could contradict. As I argued in the companion post on consumer choice, stubborn loyalty to a gatekeeper’s product is read not as a verdict to respect, but as a wall yet to be torn down. 

The second idea explains why the review hears the costs it does and misses the ones it does not. The DMA equates business-user grievances with the public interest. Its objectives are framed around benefits to business users—overcoming barriers to entry and redressing the “imbalance” between gatekeepers and their business users—and its enforcement metrics ask whether those imbalances persist.

That is a structural choice. It erases a distinction competition law has worked hard to preserve: the difference between harm to competitors and harm to competition. Once business-user satisfaction becomes a proxy for the public interest, the review will systematically overweight the voices of organized complainants who want compliance never to be declared complete. It will underweight the millions of consumers who benefit from integrated products and have no standing committee at the Commission’s compliance workshops. The feedback the regime is wired to receive is precisely the feedback that confirms it. 

Put the two together, and the result is a regulation impervious to disconfirmation. Unfairness is structural, so it is always present. After all, when will rights and obligations be in perfect balance? Exactly: never. Gatekeepers are always the beneficiaries of a lopsided distribution, so their losses never count as costs. And because unfairness is structural and axiomatic, there is no counterfactual against which to judge whether the distribution of rights and obligations is “fair.” Compared to what?

Given the conflation of public and private interests, the public interest is likely to become whatever organized business users say it is. Given the assumption that any obligation imposed by the DMA is achievable, any failure to reach that benchmark will be read as either an infringement or malicious compliance. 

A regime constituted on those premises is not self-evidently “working” just because its review says so, any more than a bus headed for a cliff is working because it is still moving. As I have repeatedly stated in tweets I now cannot unearth: Movement is not progress. 

The One-Way Ratchet

None of this would be alarming if the DMA were cheap to get wrong. It is not. The conduct it prohibits per se is precisely the category for which the error-cost literature counsels caution, because the welfare effects are context-dependent and false positives are difficult to reverse. A wrongly acquitted monopolist may still be disciplined by entry. A wrongly condemned product design is abandoned, and the foregone innovation is never observed. 

A regime that cannot perceive its own false positives will keep producing them. Worse, it will mistake the absence of complaint from the people it has overridden for evidence of success. 

The Commission describes the DMA as an “iterative process,” a reassuring phrase that implies learning. But iteration only produces learning if the loop is closed—if some outcome can send the signal: stop, reverse, this was wrong. 

The DMA’s loop runs in one direction. Each iteration adds obligations, monitors compliance, and treats shortfalls as grounds for more intervention. The first review was supposed to be the moment when the architecture faced an external test. Instead, the enforcer reviewed its own enforcement, declared the results positive, and concluded that what the regime needs is not greater scrutiny, but more vigorous application of the same approach—with cloud computing and artificial intelligence next in line. 

That is the real significance of the review. An instrument confident enough to expand before demonstrating that it can succeed is not being disciplined by evidence. It is being driven by an axiom. 

And the question the DMA’s evaluative architecture is structurally incapable of asking—whether the costs of intervention exceed its benefits—remains unanswered because it cannot be asked. The possibility that the premise itself is wrong has been designed out of the system. 

A review that cannot find error is not evidence that the policy works. It is evidence that the policy has stopped looking.

The post Fit for Purpose, Unfit for Review: Why the DMA’s First Evaluation Could Never Find Failure appeared first on Truth on the Market.

Last week, a federal court in São Paulo suspended the daily fine that Brazil’s Administrative Council for Economic Defense (CADE) had imposed on Meta for refusing to open WhatsApp to rival AI chatbots. On June 5, Turkey’s competition authority announced both an abuse-of-dominance investigation into the same conduct and an interim measure giving Meta one month to admit third-party AI assistants. Three days later, the Italian Competition Authority (AGCM)—the first agency anywhere to order interim relief against the WhatsApp restrictions—closed its case in deference to the European Commission, which had expanded its own proceedings to cover Italy. Then, on June 9, the Commission adopted interim measures of its own, giving Meta five working days to restore rival assistants’ access across the European Economic Area (EEA). Africa’s Common Market for Eastern and Southern Africa (COMESA) Competition Commission, for good measure, is investigating, too. 

Meta is not the only company in regulators’ crosshairs. At its developer conference on June 8, Apple announced that its new Siri AI features will not launch in the European Union with iOS 27, citing the Digital Markets Act’s (DMA) interoperability requirements. The next day, the Commission reportedly rejected Apple’s request for an 18-month exemption, characterizing the company’s decision as a business choice. Earlier this month, meanwhile, the UK Competition and Markets Authority (CMA) imposed its first AI-related conduct requirement on Google under Britain’s new digital-markets regime, governing how publisher content may be used in AI Overviews. 

One week. A half-dozen authorities. Four continents. 

And a common thread: nearly all of this activity is occurring before any agency has demonstrated, in a final appealable decision, that the challenged conduct actually harms competition. Interim measures, preventive suspensions, and ex ante mandates have become the enforcement tools of choice in AI markets. Whatever else one makes of these interventions, they share a defining feature: they front-load the costs of being wrong. 

The Error-Cost Framework Strikes Back

The heightened scrutiny is understandable. AI technologies are likely to become important inputs across much of the economy, making concerns about future market power inevitable. Yet competition policy risks repeating mistakes from earlier periods of technological change. Regulators may be intervening before they understand how competition in AI markets actually works.

The error-cost framework associated with the Chicago School, and most prominently with Judge Frank Easterbrook’s seminal article “The Limits of Antitrust,” emphasizes the dangers of premature intervention under conditions of uncertainty. Because markets are dynamic and often self-correcting, conduct that initially appears exclusionary may ultimately prove efficient, innovative, or beneficial to consumers. Easterbrook therefore warned that false positives—mistakenly condemning procompetitive conduct—can be especially costly.

His prescription was not inaction. Rather, it was a set of sequential filters—market power first, followed by a plausible profit-from-harm theory—to screen out cases where the expected costs of error exceed the expected benefits of intervention.

Once antitrust law prohibits a beneficial business practice, the resulting loss of innovation and competition can persist for years. By contrast, many instances of market power are disciplined over time by entry, innovation, and technological change. The asymmetry is institutional. Erroneous condemnations become embedded in precedent and, if corrected at all, are usually corrected only through slow appellate or legislative processes. Erroneous acquittals remain subject to market discipline.

Put differently, Type II errors come with a built-in correction mechanism. Type I errors do not. The lesson is not that antitrust should never intervene. It is that intervention warrants particular caution when regulators have limited confidence about how markets will develop.

Many contemporary policy discussions implicitly assume that AI markets are becoming concentrated and therefore require early intervention. Current market realities suggest a far more fluid competitive environment. 

OpenAI, Anthropic, Google, xAI, Meta, DeepSeek, Alibaba, Mistral, and numerous other firms continue to compete intensely across multiple dimensions, including model performance, cost, deployment strategies, and ecosystem development. Leadership at the technological frontier has changed hands repeatedly over the past three years. The DeepSeek episode—in which a comparatively small lab matched state-of-the-art performance at a fraction of the presumed cost—should have put to rest the notion that today’s leaders are already entrenched, at least for now. 

Ironically, the partnerships between cloud-computing incumbents and AI startups that most concern regulators have arguably financed much of this competitive entry. Notably, every merger authority to examine those partnerships reached the same conclusion. The CMA closed all five of its AI-partnership reviews without remedies. After a 15-month investigation of Microsoft/OpenAI, the agency concluded that even the deepest of those relationships fell short of conferring control. 

Competition also increasingly takes place across several interconnected layers of the AI stack, including semiconductors, cloud infrastructure, foundation models, application-layer services, and deployment environments. Success in one layer does not necessarily translate into durable dominance in another. As one of us has argued at length elsewhere, the most important competitive bottlenecks in AI may emerge at deployment interfaces—the points where users and businesses actually access AI services—rather than within foundation models themselves. 

This dynamism should give regulators pause before assuming competitive outcomes that have yet to emerge. 

Maybe Data Moats Aren’t Moats After All

A second feature of AI markets deserves attention, though it warrants more caution. 

It is increasingly claimed—by both AI enthusiasts and critics—that AI systems improve through continuous deployment and feedback. The claim is most plausible in robotics and other physical AI applications. Unlike traditional software, which is updated periodically, these systems may evolve through ongoing interactions with users, environments, sensors, and operational constraints. 

Industrial robots, autonomous systems, logistics platforms, and enterprise AI applications generate enormous amounts of operational feedback. On this account, that feedback becomes an input into future improvements. Learning is not confined to the development stage but continues after deployment. Japan’s Fair Trade Commission (JFTC) devoted a new chapter of its generative-AI market study to autonomous driving this spring for precisely this reason. 

The result, at least in theory, is a recursive learning process. Deployment generates data. Data improves performance. Improved performance encourages further deployment. Additional deployment creates new learning opportunities.

The key question is how powerful these feedback loops actually are. Having enough data typically matters far more than having the most, and the returns to additional data often diminish rapidly. The same dystopian narrative that once predicted insurmountable data moats in search and social media is now being replayed for foundation models. Whether AI deployment loops will prove more durable or consequential than their search-engine predecessors remains an open empirical question, not a premise to be assumed.  

Yet these feedback loops sit at the heart of virtually every theory of harm currently advanced in AI markets. The Federal Trade Commission’s (FTC) Section 6(b) report on AI partnerships relies heavily on them. So do the remedies in the U.S. Google search case, which were expressly extended to generative AI to prevent Google from converting purported data advantages into AI dominance. 

Authorities that invoke these loops to justify intervention must also accept the corollary. If the loops are real and economically significant, interventions that sever them may destroy far more value than static analysis suggests. If the loops are weak, many of the underlying foreclosure theories collapse on their own terms. Either way, the case for aggressive early intervention appears weaker than it first seems. 

The Cost of Being Early

Traditional antitrust interventions often assume relatively stable market structures and that corrective measures can be imposed without significantly affecting innovation. AI markets may be different. 

If technological progress depends on recursive learning processes, premature intervention may disrupt the very mechanisms through which innovation occurs. Restrictions on product integration, limits on deployment strategies, or mandatory redesigns imposed before competitive harm is established may reduce experimentation, slow learning, and ultimately diminish innovation. 

The costs of false positives would then extend beyond conventional efficiency losses. They could alter the trajectory of technological development itself. 

That possibility makes the current enthusiasm for interim measures especially striking. Under Article 8 of Regulation 1/2003, interim relief requires only a prima facie finding of infringement and a risk of serious and irreparable harm to competition—a standard so exceptional that the European Commission invoked it only once during the regulation’s first two decades. By design, it is the enforcement tool most exposed to error costs because it rests on predictions rather than demonstrated effects. 

Yet in the WhatsApp saga, interim measures have become the tool of choice. Italy used them in December 2025. Brazil followed in January 2026. Turkey and the Commission acted this month. Authorities across four continents are now rewriting the terms on which a private messaging platform deals with AI developers in real time, before any of them has established that consumers were harmed. 

The concern becomes even more significant in physical AI systems. Robotics, autonomous vehicles, industrial automation, and other deployment-intensive applications depend on continuous real-world experimentation. Learning cannot be fully replicated in laboratories or regulatory sandboxes. It emerges through actual operation. 

A mistaken intervention that disrupts those learning processes may therefore have consequences that persist long after the underlying enforcement error becomes apparent. 

Gatekeepers, Special Responsibilities, and Other Status Crimes

These concerns are particularly relevant to the growing tendency to apply doctrines that focus more on firm status than demonstrated competitive effects. 

One example is the European concept of “special responsibility,” first articulated in Michelin v Commission and steadily expanded since. Under the doctrine, firms deemed dominant often face obligations that go beyond ordinary competition-law standards. 

When detached from a concrete assessment of exclusionary effects, special responsibility begins to function as a status-based constraint on conduct. Competition authorities and courts may come to treat product integration, ecosystem design, or self-preferencing as presumptively suspect simply because a successful dominant firm engages in them. That approach departs from competition-on-the-merits principles. 

Nor is this logic confined to Article 102 case law—or to Europe. It is hard-coded into the DMA, whose obligations attach to designated gatekeepers regardless of demonstrated effects. The result is the spectacle of European consumers losing access to Siri AI while Brussels and Cupertino argue over whose “choice” that was. 

The same logic animates the Commission’s pending Article 6(7) specification proceedings, which concern how Google must provide rival AI services with “equally effective” access to Android’s hardware and software features. One of us has criticized that interoperability theory as more slogan than substance. Similar thinking underlies the CMA’s new conduct requirements under the United Kingdom’s strategic-market-status regime. It also echoes through the WhatsApp interventions in Brazil and Turkey. 

AI may therefore become the first major technology to be born into a status-based regulatory environment, rather than growing up under effects-based antitrust scrutiny. 

The same impulse is beginning to reshape adjacent doctrines. Last month, Brazil’s CADE ordered the retroactive notification of Microsoft’s Inflection acqui-hire, a transaction that fell below every applicable filing threshold, on the theory that talent and licenses can substitute for an acquisition of control. The agency simultaneously opened new probes into Google’s Windsurf and Hume AI arrangements. Whatever the merits of treating acqui-hires as concentrations, the direction of travel is unmistakable: jurisdictional doctrines are being remodeled to reach AI transactions before their effects can be observed. 

This danger is particularly acute in AI markets. Product integration is often not a strategy for exclusion, but a mechanism for improving performance and, if the deployment-learning thesis is correct, accelerating learning. 

Integrating AI assistants into operating systems, embedding models within enterprise software, or optimizing interactions between hardware and software may generate substantial efficiencies. Those efficiencies often arise precisely because integration enables faster feedback, tighter coordination, and more effective learning. 

Condemning such conduct as self-preferencing without a concrete demonstration of exclusionary effects may amount to penalizing successful innovation rather than protecting competition. Treating integration as unlawful merely because it advantages a firm’s own products departs from competition-on-the-merits principles and risks undermining the very innovation process that competition policy is supposed to protect. 

Back to First Principles

None of this implies that AI markets should be exempt from antitrust scrutiny. 

Exclusionary conduct remains possible. Firms that control critical infrastructure, computing resources, data assets, or distribution channels may engage in practices that genuinely foreclose rivals. If durable market power emerges and is used to restrict competition, antitrust law should respond. 

The point is not that intervention is never warranted. It is that intervention should remain grounded in evidence of competitive harm, rather than assumptions based on firm size, ecosystem integration, or speculative fears of future dominance. 

Nor is such restraint merely theoretical. In the U.S. Google search case, Judge Amit Mehta rejected both structural divestiture and a proposed requirement that Google provide advance notice of its AI investments beyond existing Hart-Scott-Rodino obligations because the evidentiary record could not support either remedy. Taiwan’s Fair Trade Commission reached a similar conclusion from the opposite direction. After a year of consultation, it determined that generative AI requires no new enforcement instruments. The existing Fair Trade Act, applied through an issue-driven rule-of-reason framework, is sufficient. 

Effects-based enforcement is not a euphemism for inaction. It is a methodological commitment to identifying actual competitive harm before imposing remedies. 

The fundamental question remains the same one that has guided sound antitrust policy for decades: does the conduct harm competition, or merely competitors? 

The Machines May Be Learning. Regulators Should, Too.

The emergence of AI does not undermine the traditional error-cost framework. If anything, it reinforces its central insight. AI markets are characterized by extraordinary uncertainty, rapid technological change, and intense competition among alternative technological pathways. Under such conditions, regulators are especially likely to mistake efficient conduct for anticompetitive conduct. 

That danger is magnified to the extent that AI systems—as many enforcement theories themselves assume—improve through recursive learning, deployment-driven experimentation, and continuous feedback from real-world use. Interventions that disrupt those processes may do more than reduce static efficiency. They may alter the trajectory of innovation itself. 

In AI markets, the costs of false positives therefore extend beyond the immediate consequences of a particular enforcement decision. They may affect the pace and direction of technological development for years to come. 

For that reason, the error-cost framework deserves renewed attention. Far from rendering Judge Easterbrook’s insights obsolete, AI may strengthen the case for caution when competitive effects remain uncertain. 

The events of the past week suggest that enforcement is moving well ahead of that insight. Competition authorities should remain vigilant against genuine exclusionary conduct. They should also recognize that preserving experimentation, deployment, and innovation is itself a core objective of competition policy. 

In markets defined by continuous learning and technological change, the competitive process is often best protected through patient, effects-based enforcement rather than remedies imposed before the evidence is in. 

After all, if AI’s defining feature is that it learns from experience, competition policy should do the same.

The post Act First, Learn Later: AI Antitrust and the Error Costs of Regulation at Machine Speed appeared first on Truth on the Market.

Late last month, the California Assembly passed the COMPETE Act, based on recommendations from the California Law Revision Commission. The bill is now under consideration in the California Senate. The act would significantly expand liability under the Cartwright Act, California’s antitrust statute, to such an extent that it would likely become the most interventionist state antitrust law in the country.

As currently drafted, the statute could place a wide range of common business practices at legal risk. That would make California an antitrust outlier and could discourage investment and economic growth. At the same time, there is no assurance the statute would benefit consumers and, in some circumstances, it could even leave them worse off.

Antitrust by Guesswork

The statute’s key feature is a new cause of action that would allow suits against businesses for anticompetitive conduct by a single firm, as distinct from collusion among multiple firms, which the law already covers. The federal Sherman Act also provides a single-firm cause of action, but the California statute makes a critical change.

The statute would make firms liable for conduct that “monopolizes” or “unreasonably restrains trade,” rather than only conduct that “monopolize[s],” as the federal statute provides. That added phrase makes all the difference.

Monopolization suits inherently require courts to identify exclusionary conduct without punishing business practices that are simply part of vigorous competition. Over more than a century, federal courts have developed—through trial and error—a structured framework to meet that challenge. By contrast, there is no settled understanding of what it means to “unreasonably restrain trade” in the single-firm context.

Businesses therefore would have little guidance on compliance, especially because the statute also provides that courts should treat federal antitrust precedent as “at most, instructive.” Guess wrong, and the result could be treble damages or civil or criminal penalties, which were increased by legislation enacted in December 2025. While the statute excludes independently owned, California-based businesses with $10 million or less in annual “gross receipts” and 100 or fewer employees, it still leaves a wide range of medium and large businesses—including franchisees of large national businesses—exposed to significant liability.

The risks of overenforcement are compounded by the statute’s instruction that courts interpret the Cartwright Act “liberally” to “maximiz[e]” effective deterrence. That contrasts with the more balanced approach in federal antitrust case law, which strives to deter anticompetitive conduct without suppressing procompetitive practices.

The statute also pursues a wide range of policy objectives, such as “free and fair competition,” protecting “all trade participants,” and preserving “democratic, political, and social” institutions. That broad remit strains the rule of law and could yield enforcement actions and judicial decisions that depart from antitrust’s traditional focus on preserving competition and protecting consumers.

Taking the Guardrails Off

The risk of enforcement overreach is heightened by the statute’s unusually specific instructions that courts need not follow several doctrines developed in federal monopolization cases. These doctrines share a common purpose: reducing the risk that antitrust law mistakenly condemns procompetitive business practices. The result could be enforcement outcomes that protect certain competitors at consumers’ expense.

Predatory Pricing. The statute disclaims Supreme Court precedent—most notably, Brooke Group Ltd. v. Brown & Williamson Tobacco Corp.—requiring predatory-pricing claims to show both below-cost pricing and a reasonable prospect that the defendant can later recoup its losses through monopoly pricing. Courts developed these requirements to avoid discouraging price cuts that benefit consumers. Without them, businesses unable to match a rival’s prices could seek relief through antitrust litigation instead.

Multi-Sided Platform Efficiencies. The statute disclaims Supreme Court precedent—specifically, Ohio et al. v. American Express Co. (2018)—requiring courts to consider whether alleged anticompetitive harms on one side of a multisided platform are offset by procompetitive benefits on another side. A multisided platform is a business that serves distinct groups of users who interact through the platform, such as advertisers and consumers using a search engine. Without that framework, a court could condemn certain practices in digital advertising markets without considering how those same practices may benefit consumers through zero-price search, messaging, or other complementary services.

The “As Efficient” Principle. The statute disclaims certain federal appellate decisions requiring plaintiffs in monopolization cases—particularly those involving bundled discounts—to show that the challenged conduct impedes rivals that are at least as efficient as the defendant. The principle reflects the view that monopolization law should not protect less efficient competitors from more efficient ones. Abandoning it could discourage practices such as bundled discounts and other package deals that often benefit consumers.

If You Can’t Win in Court…

Viewed more broadly, the statute’s interventionist approach reflects a shift from the federal to the state level of a vigorous campaign—in scholarship, advocacy, and policymaking—to move U.S. antitrust law away from protecting the competitive process and toward achieving redistributive outcomes by limiting firm size and market concentration as ends in themselves. As I have shown elsewhere, federal courts have largely rejected legal theories motivated by these redistributive principles, which generally conflict with governing case law grounded in the consumer-welfare standard and the effects-based rule-of-reason methodology. Advocates have therefore turned to state legislatures to implement this alternative model of antitrust law.

Supporters of this state legislation rely heavily on the now-popular view that purportedly lax antitrust enforcement has increased market concentration and entrenched incumbents. Yet the real world is more complicated.

There is no scholarly consensus that the U.S. economy is unusually concentrated by historical standards or that antitrust policy was the determining factor in those digital market segments where concentration is high. Sectors with high concentration can still feature intense competition on price, quality, and innovation, as the current rivalry among major cloud-computing services and artificial-intelligence model developers illustrates. Moreover, it is well-established that size and concentration within a certain range can benefit consumers when economies of scale or scope—cost savings or product improvements that come from producing more goods or offering complementary services—yield gains in cost, quality, or convenience that are passed on to consumers.

The pending statute departs from core principles of antitrust law and economics to address a competition problem that has not been soundly demonstrated. Its loosely defined cause of action, expansion of policy goals beyond protecting competition, and rejection of doctrinal guardrails against overenforcement may encourage litigation targeting practices that pose no reasonable risk of anticompetitive harm. The statute arguably goes beyond even the broader liability standards of European competition law, which antitrust reformers often laud as a model, because it is enforceable through class actions and treble damages—plus attorney’s fees—that are uncommon in Europe.

The Golden State’s Self-Inflicted Wound

The combination of expansive liability standards and private enforcement could significantly alter California’s legal environment for the medium and large businesses—and small franchisees—that fall within the statute’s scope. Capital is mobile, and competitive pressures require firms to avoid regulatory environments that place them at a disadvantage.

If enacted in its current form, the statute may lead national businesses to shift investment and employment to other states. It may also discourage discount pricing, package deals, and other business practices that benefit California consumers—especially lower-income consumers—but cannot be matched by less efficient rivals.

In the name of protecting competition, California risks making itself less competitive. The ultimate losers may be the very consumers antitrust law is supposed to protect.

The post California’s Other Wealth Tax appeared first on Truth on the Market.

When Mozilla reports that Firefox is now selected through a DMA browser-choice screen once every 10 seconds—more than 6 million selections in total, with daily active iOS users 113% higher in the European Union than they would have been absent the regulation—the result is treated as proof that the DMA is working. Consumer choice has spoken. 

When fewer than 1% of Meta users chose its new paid, ad-free subscription option—introduced largely to comply with DMA edicts—over the existing free, ad-supported alternative, that result was treated as proof that the DMA needs to go even further. Critics pointed to the low uptake as evidence that Meta had priced the subscription to deny users a “genuine” choice. Consumer choice, it seems, needed a chaperone. 

Notice the asymmetry. In the first case, consumer choice is a verdict to be respected. In the second, it is a symptom to be cured. The common thread is not consumer welfare, but direction of travel. Choice matters when it moves users away from a designated gatekeeper. When it does not, regulators and commentators look for reasons to discount it. 

I develop this argument at greater length in a new International Center for Law & Economics (ICLE) white paper. It nonetheless merits separate treatment because it captures something about the DMA that more familiar critiques—focused on per se rules, limited judicial review, or the absence of an effects-based analysis—do not fully reach. 

The DMA does not ignore consumer choice. It instrumentalizes it. Choice enters the analysis on one condition: it must support more intervention. 

Choice for Me, but Not for Thee

Competition law, for all its faults, treats consumer behavior as informative. The entire consumer-welfare framework rests on the premise that what consumers actually do—what they buy, switch to, or stick with—tells us something about whether conduct is helping or harming them. 

If consumers continue choosing an integrated product after being shown alternatives, that is evidence. It may suggest the product is better, cheaper, or more convenient. It may suggest switching costs are real, but that the benefits of switching are modest. Effects analysis exists precisely to sort among those possibilities, rather than assume the answer in advance. 

The DMA dispenses with that inquiry. Its objectives of “fairness” and “contestability” are defined, by the Act’s own terms, independently of “the actual, potential or presumed effects of the conduct of a given gatekeeper … on competition on a given market.” In other words, the DMA’s core concepts do not depend on showing that conduct harms competition or consumers. 

Recital 33 defines fairness only in negative terms, with “unfairness” deemed to be “an imbalance between the rights and obligations of business users where the gatekeeper obtains a disproportionate advantage.” There is no benchmark, no counterfactual, and—crucially—no role for revealed preference. What consumers actually choose places no meaningful constraint on intervention. 

What emerges is a regime with a built-in direction, but no clear destination. The premise of structural unfairness is that designated markets produce unfair outcomes because they are structurally unfair. Whenever an asymmetry remains—whenever users have not migrated, whenever a gatekeeper’s product remains popular—the diagnosis is the same: not enough has been done yet. 

The logic is self-reinforcing. Persistent gatekeeper success is not considered evidence that might falsify the premise. It is taken as evidence that the premise has not yet been fully implemented. 

That helps explain why the DMA struggles to process the most mundane explanation for why people keep using gatekeepers’ products: fully informed consumers may simply prefer them. A framework that took that possibility seriously would have to treat sustained uptake of an integrated product as a reason to stop intervening—or at least to reconsider whether the intervention is working as intended. The DMA is structurally inclined to reach the opposite conclusion. Heads, the regulator wins; tails, the consumer switches. 

Fairness Isn’t Free

None of this would matter much if intervention were costless. It is not. Reshaping how consumers interact with core platform services is not a side effect of the DMA. It is the mechanism. And that reshaping carries costs that fall, at least initially, on the very consumers in whose name the regime ostensibly operates.

Some of those costs are already visible. Mozilla’s browser-choice data came with an acknowledgment that users now face more friction when navigating online. Apple has delayed or withheld a series of features in the European Union—including iPhone Mirroring, Live Translation for AirPods, and, most recently, its revamped Siri AI—each time citing the difficulty of complying with interoperability mandates without exposing user data.  

One can debate, as the European Commission and Apple do, how much of this reflects genuine engineering constraints and how much amounts to strategic foot-dragging. The debate itself nonetheless concedes the point. The regime is changing the consumer product and, at least in some cases, making it less convenient or less capable. 

There is a respectable case for accepting those costs. A DMA defender can argue openly that some consumer convenience, quality, and integration are worth sacrificing to achieve less concentrated markets and greater competition over time. Some proponents say exactly that: short-term inconvenience is the price of long-term contestability. That is a coherent position. 

It is also an inherently political one. It reflects a choice to benefit one set of interests—organized business users, rivals, and complementors—at some expense to another: consumers who prefer the status quo. 

The trouble is that the DMA rarely presents the tradeoff in those terms. Instead, it is packaged in the universal language of “fairness,” a word that suggests nobody loses. 

Yet the Organisation for Economic Co-operation and Development (OECD) has acknowledged that fairness in this context is “strongly tied to redistribution.” The Commission’s own expert panel similarly described the fairness objective as achieving “a higher standard of fairness in the distribution of the social value generated by large platforms.” 

Redistribution may be a defensible objective. But redistribution from consumers to business users, carried out in the name of consumers, is something else entirely. 

My colleagues and I have made a version of this argument before. The Google and Apple designation decisions illustrate how little users actually matter under the DMA, even as policymakers invoke their interests at every turn. 

The Difference Between Policing and Steering

That sleight of hand is easier to see once we are precise about the kind of regime the DMA creates. The problem is not merely that the DMA sometimes discounts consumer preferences or understates consumer costs. It is that the DMA treats consumer behavior differently from competition law because it is doing a different job. 

The DMA is not simply an “oversight” regime. It is an “ordering” regime. And that distinction is the whole ballgame. 

An oversight regime—such as competition law in its Article 102 Treaty on the Functioning of the European Union (TFEU) form—polices markets from the outside. It treats consumer behavior as a partly exogenous input: something to be measured, explained, and respected, even when the explanation is inconvenient.

If consumers keep buying the integrated product, the enforcer must grapple with that fact. The burden is to show harm to consumers, and persistent consumer uptake is evidence that demands an explanation. A diffuse and numerous consumer class cannot easily be conscripted into any particular firm’s commercial agenda. 

An ordering regime works in the opposite direction. It begins with a vision of what the market should look like—more rivals, more switching, more disaggregation—and treats consumer behavior as an outcome to be steered toward that vision. 

Choice architecture ceases to be something regulators should approach neutrally and becomes a policy tool. The question is no longer, “What do consumers reveal that they prefer?” It becomes, “Are consumers moving in the right direction, and if not, what additional steering is required?” 

Consumer choice is not ignored. It is repurposed—from a constraint on the regulator to an instrument of the regulator. 

That repurposing helps explain why the regime is resistant to its own failure. Every outcome can be interpreted as validation. Significant migration to rivals vindicates the intervention. Persistent loyalty to gatekeepers justifies intensifying it. 

The European Commission’s first review of the DMA reflects precisely this disposition. It treats designations, proceedings, and enforcement outputs as evidence of success, while treating reports of consumer harm as consultation inputs rather than findings that might call the underlying premise into question. 

A bus headed for a cliff is not “working” simply because it is moving. 

What If They Already Chose?

There is an older intellectual lineage here, though not the one usually invoked in the DMA’s defense. 

Ordoliberalism, the tradition that shaped postwar European competition governance, is generally remembered for its commitment to an “economic constitution“: a rules-based order insulated from day-to-day political pressure and policed by a neutral guardian, rather than steered toward preferred outcomes. On that telling, ordoliberalism is the opposite of market ordering. 

But the tradition always carried a more directive strand. Franz Böhm, one of ordoliberalism’s founders, described the “true core of an economic constitution” as “a steering norm, which guides economic happenings in a politically desirable direction.” He also accepted that the state should suspend “indirect economic steering by legally-governed competition” whenever “direct market steering by methodical means of command” seemed more appropriate. 

The competitive order was to be tended by a nobilitas naturalis—an “aristocracy of the public spirit” trusted to read political needs and pull the right lever. 

The DMA is that directive strand, shorn of the constitutional constraints that were supposed to discipline it. The European Commission is cast as the aristocracy of the public spirit, charged not with refereeing the competitive process, but with deciding what “fair” digital markets should look like and steering consumers toward them. 

Once the regulator’s role is to define desirable outcomes rather than police a competitive process, consumer choice loses its standing as an independent fact. It becomes one more input for the steering norm to interpret—respected when it cooperates, corrected when it does not. 

That is the deepest sense in which the DMA instrumentalizes choice. The problem is not merely that the Commission reads the data tendentiously. Under an ordering logic, there is no other way to read it. A regime built to redesign markets around a preferred template cannot also treat consumers’ revealed preferences as a verdict on whether that template was worth pursuing. It can only treat them as a progress report. 

That leaves the question the DMA is structurally unequipped to ask: What if consumers do not need steering? What if they already chose? 

The post If at First Consumers Don’t Switch, Regulate Again appeared first on Truth on the Market.

That day, the U.S. government cut non-Americans off from access to two of the world’s most advanced AI models. Brussels’ flagship response to precisely this scenario—the Cloud and AI Development Act (CADA)—would not have helped. 

As I argue in “Europe’s Sovereignty Stack: CADA, Compute, and the Limits of Autarky,” Europe needs to do two things instead: remove legal barriers and build serious computing capacity on the continent. That strategy would allow Europe to pursue two complementary goals. 

First, it would give Europe leverage through mutual dependence, strengthening its hand in negotiations over access to cutting-edge AI capabilities that it has little realistic chance of developing domestically in the near term. Second, it would reduce dependence on foreign providers for “good enough” AI systems, which will remain economically important across a wide range of applications that do not require state-of-the-art models. 

The Alliance, the Act, and the Kill Switch

A U.S. export-control directive ordered Anthropic, a leading American AI lab, to suspend non-U.S. access to its two most capable models—Fable 5 and Mythos 5—after a jailbreak exposed a security vulnerability. The company complied within hours, pulling the models globally for non-U.S. users while disputing the order’s proportionality. Its other models remained available. 

That was not an isolated event. It followed a remarkable 10-day run. 

On June 2, the White House issued an executive order granting the federal government up to 30 days of pre-release access to “covered frontier models”—advanced AI systems at the edge of current capability—and a role in choosing the “trusted partners” who get them first. 

On June 3, the European Commission proposed CADA, the centerpiece of its Technological Sovereignty Package. 

That same day, EU member states moved to join “Pax Silica,” the U.S.-led chip alliance, under an understanding to buy at least $40 billion in American AI chips. 

So the European Union decided to bind itself more tightly to the American stack—the layers of chips, cloud infrastructure, and models that make modern AI possible—while CADA proposed walling its most sensitive workloads off from that same stack. Then the kill switch fired. 

Open to the World, Closed to the Providers

CADA would create four “assurance levels” for public-sector cloud and AI services. The sensible part is that the framework is risk-based and applies only to the public sector. 

At the highest assurance levels, though, CADA seeks to shield European buyers from third-country interference by effectively excluding even American-owned providers. 

Henna Virkkunen, the European Commission vice president responsible for technology sovereignty, made that goal explicit. She told reporters that no provider handling critical workloads should have a “kill switch,” that “we want to make sure that our most critical sensitive data is stored in Europe,” and, according to press accounts, that U.S. firms would struggle to qualify for the highest sovereignty tier because of the U.S. CLOUD Act. 

CADA contains a single statutory escape hatch from its ownership restrictions, and it does not apply at the highest assurance level. A provider’s home country can be recognized as an “associated third country,” but the criteria appear tailor-made to exclude the United States. One provision would disqualify countries that maintain measures that “impede the provision of state-of-the-art technologies.” 

Read literally, that would disqualify a country for imposing the very export controls that underpin the Western technology alliance. It would exclude a partner for participating in the same system the European Union joined through Pax Silica during the very week CADA was unveiled. 

In a communication accompanying the proposal, the Commission insisted that technological sovereignty “does not mean isolation, protectionism, or tech decoupling,” and promised that the European Union would remain “open to the world.” That assurance is difficult to reconcile with both the text of the proposal and Commissioner Virkkunen’s public statements. 

The Problem Isn’t Dependence. It’s Leverage.

Perhaps the biggest flaw in CADA’s approach is that it effectively gives up access to the best AI capabilities precisely where they matter most: defense, intelligence, and other national-security applications. 

To be sure, CADA includes a safety valve that permits exceptional purchases when “no adequate or reasonable alternative” exists. But the exemption comes with conditions—such as prior unsuccessful tenders—and appears deliberately designed to be difficult to use in practice. That is a curious approach given the Commission’s own analysis, which acknowledged that some degree of dependence is “inevitable.” 

No amount of wishful thinking—or attempts to redefine the term—can change the reality that frontier AI is available only from the leading U.S. labs. Europe cannot develop comparable capabilities quickly enough to capture both the economic and military-intelligence benefits of being at the frontier. 

In highly competitive domains such as cybersecurity, even a delay of a few months can matter. Nor is there an obvious alternative source. The strongest non-U.S. AI labs are largely based in China, not Europe, and most assessments place them at least several months behind the leading American firms, with larger gaps in some areas. 

The June 12 export restriction exposed a deeper problem: the European Union has little leverage when it comes to negotiating access to state-of-the-art AI capabilities. Building that leverage—not restricting access to the technologies Europe already lacks—should be the real focus of European policy. 

Build the Bargaining Chip, Not the Fantasy

The immediate answer is not a sovereign European frontier model. That capability is not arriving on any meaningful timeline, and focusing exclusively—or even primarily—on that goal would waste scarce time and political attention. 

The better answer is to build the one thing Europe realistically can: compute capacity on European soil, measured in tens of gigawatts. What matters most is not who owns the accelerators or whose models run on them. It is where the metal sits. 

European compute could provide the leverage Europe lacks. Demand for computing capacity will outstrip even American supply. A European Union with large, modern data-center capacity to offer—and a willingness to host U.S. firms’ workloads—would create the kind of mutual dependence that makes a future cutoff too costly to contemplate. If U.S. firms run not only European workloads but also U.S. workloads in European data centers, then cutting off Europe would not merely sacrifice non-U.S. revenue. It would directly harm U.S. customers. 

To build that capacity—and win enough of the market to create real dependence—Europe needs to move fast. It needs permitting, including environmental permitting, and grid connections quick enough to make investing in the European Union more attractive than waiting elsewhere. In that legal environment, public subsidies may not be necessary. Without those changes, even large subsidies will not be enough. 

On security, Europe should draw the line along the right axis. Instead of treating any third-country control as suspect, CADA should target adversary control. And the recognition decision should become an openly political judgment about allies and adversaries, not a legalistic trapdoor for excluding partners. 

The main security burden should fall on measures that depend far less on the provider’s nationality: customer-held encryption keys, protection for data while it is being processed, vetted and security-cleared staff, anti-tamper controls, and portability. 

Bolted to Our Floor

In CADA, the Commission got one thing right: the need to accelerate permitting for data centers and related infrastructure. It just was not nearly ambitious enough. The proposal pairs a modest fast-track with promises of insulation from foreign control that its own impact assessment all but concedes are unattainable, given Europe’s dependence on American frontier AI. 

The report “Europe 2031,” written by researchers who want a more competitive Europe, sketches where a CADA-like strategy could lead. Public institutions most committed to “Buy European” become “the ones paying the ransoms,” outmatched by competitors using better foreign tools. The compute gap never closes. Washington rations access to frontier AI by alliance tier. Economic and political pressures mount as Europe finds itself with fewer options and less leverage. 

The report’s prescription is simple: make sure the data centers that get built are “bolted to our floor.” 

That is the real lesson of June 12. Europe does not need the illusion of autonomy. It needs leverage. And leverage comes not from pretending dependence can be wished away, but from building assets others cannot afford to lose.

The post Bolted to the Floor, Not Written in the Clouds appeared first on Truth on the Market.

That is the problem at the center of the current debate over free speech in the digital age. For years, the public argument focused on whether private technology companies were moderating too much content. A different concern has now moved to the forefront: “jawboning,” the practice of government officials using informal pressure, implicit threats, or regulatory leverage to induce private intermediaries to suppress speech on the government’s behalf.

Late last week, Sens. Ted Cruz (R-Texas) and Ron Wyden (D-Ore.) introduced the Justice Against Weaponized Bureaucratic Overreach to Networked Expression Act, or JAWBONE Act. The bipartisan legislation seeks to bring jawboning into the open and provide Americans with a legal mechanism to challenge government efforts to coerce private actors into censoring lawful speech. 

The proposal has much to recommend it. At the same time, it should be viewed as only a first step toward curbing the federal government’s ability to conduct coercive speech-suppression campaigns. To be fully effective, it should be paired with structural reforms that limit government officials’ discretion to pressure online platforms over speech decisions. 

How to Censor Without Looking Like It

What the JAWBONE Act gets right is its recognition that modern censorship rarely takes the form of a direct government order to stop speaking. More often, it operates through censorship by proxy. 

As I argued in a recent paper on jawboning, much of today’s speech ecosystem depends on multisided platforms—businesses that connect content creators, audiences, advertisers, payment providers, and other participants. Because these platforms often depend on government licenses, regulatory approvals, or antitrust clearance, they can be particularly vulnerable to government pressure. 

When officials exert pressure on one part of that ecosystem, the effects can spread throughout the network. If the Federal Communications Commission (FCC) hints that a broadcaster’s license may be at risk, or if the Federal Trade Commission (FTC) opens an investigation into an advertising agency or fact-checking organization, the consequences can extend well beyond the immediate target. Those actions can alter incentives across the platform and, in turn, affect what speech reaches the public. 

The result is a marketplace of ideas shaped less by voluntary choices and more by the preferences of those who hold government power. Because these interventions often occur indirectly, they can evade meaningful First Amendment scrutiny and leave the public unaware of the government’s role in influencing speech outcomes. 

Everyone Loves a Shortcut

The JAWBONE Act’s bipartisan sponsorship reflects an important reality: jawboning is a tool of institutional power, not a uniquely partisan tactic. Without stronger accountability mechanisms, whichever party controls the executive branch will have incentives to use government leverage to influence speech indirectly. 

The statements Sens. Cruz and Wyden issued alongside the bill underscore this point. Republicans have argued that the Biden administration used the Cybersecurity and Infrastructure Security Agency (CISA) and other federal agencies to pressure major technology platforms to suppress content related to COVID-19 and election integrity. Democrats, meanwhile, have criticized actions by the Trump administration and FCC Chairman Brendan Carr, alleging that threats of regulatory scrutiny were used to pressure broadcasters over disfavored late-night programming and political candidates appearing on programs such as The View. 

My paper identifies additional examples. When the FTC investigates fact-checking organizations such as NewsGuard or pressures Apple over editorial decisions in news curation, the underlying mechanism resembles earlier efforts by the Obama and Biden administrations to encourage banks and payment processors to cut ties with disfavored businesses or political causes. The targets may differ, but the basic dynamic remains the same: government officials pressure private intermediaries, which then shape what speech, information, or viewpoints reach the public. 

That dynamic can produce a form of backdoor censorship that distorts the marketplace of ideas while obscuring the government’s role in the process. 

Giving the First Amendment a Remedy

The JAWBONE Act is designed to target the legal barriers that have historically allowed jawboning claims to evade judicial review. In cases like Murthy v. Missouri, plaintiffs have faced significant hurdles in proving standing, and courts often dismiss cases as moot when policies change or officials leave office. 

The bill addresses those structural barriers in several ways. First, it creates a federal cause of action that allows individual Americans to sue government agencies or employees who coerce platforms into suppressing protected speech. It also permits plaintiffs to seek monetary damages, which may help address some of the standing problems presented in Murthy. 

Second, the bill would impose liability for attempted coercion. A violation occurs regardless of whether the censorship succeeds. That feature directly addresses the chilling effect of backdoor censorship efforts: once an official exerts coercive pressure, the legal line is crossed, even if the platform refuses to comply. 

Finally, the bill would require federal agencies to submit disclosures to Congress regarding certain communications with private platforms. By bringing those communications into the open, the bill targets one of jawboning’s central advantages: secrecy. 

You Can’t Sue Away Discretion

While the JAWBONE Act would represent a major step forward for accountability, transparency, and judicial review, lawsuits alone cannot fully protect the marketplace of ideas. As long as government agencies retain broad discretion over speech intermediaries, opportunities for jawboning will remain. 

The bill includes an exception for government action “taken pursuant to a lawful investigation under, or the enforcement of, Federal or State law” that “does not violate the First Amendment to the Constitution of the United States.” That carveout may be necessary to preserve agencies’ ability to perform legitimate regulatory functions. At the same time, it creates a potentially significant loophole whose boundaries courts would need to define. 

The FCC, for example, would likely argue that investigations of broadcast licensees fall squarely within its statutory authority to determine whether broadcasters serve the “public interest.” The FTC could similarly contend that investigations of speech platforms merely reflect its responsibility to police “unfair or deceptive” practices. 

The key legal question would be whether such actions cross the line into coercion under the standard articulated in Bantam Books v. Sullivan and reaffirmed by the Supreme Court in NRA v. Vullo. As I noted previously in evaluating the Jimmy Kimmel controversy: 

The problem for Jimmy Kimmel here is that the FCC has very broad authority over broadcasters because they can review the use (and transfer) of licenses under the “public interest” standard. Caselaw suggests the FCC can revoke licenses, deny their transfer, and issue fines for things like not giving political candidates the right to respond, news distortion, and broadcasting obscenity. As will be discussed below, the Supreme Court has allowed the FCC to be the cop on the beat, so to speak, in policing whether licensed broadcasters are acting “in the public interest” with little restriction on how the FCC defines those terms.

This makes the situation very different than the Rhode Island commission [in Bantam Books], which could, at best, merely recommend prosecution for obscenity. Here, the FCC is more like the cop who is giving a warning to stop breaking the law. The import of this difference would be the source of debate in a case brought by Kimmel against the FCC…

In this case, Kimmel could argue that Carr’s statements “went far beyond advising… [on] legal rights and liability” to a “scheme of state censorship.” For its part, the FCC could respond that Carr was merely “aiding” local broadcasters to comply with the FCC’s regulations, which the FCC has authority to enforce. A court would need to determine whether the comments crossed the line into coercion aimed at censorship, or whether they were merely advising broadcasters of their responsibilities. It isn’t clear which way this would go.

That uncertainty points to a broader problem. When an agency possesses licensing authority, investigative powers, or other forms of regulatory leverage, even a polite suggestion about how the law might apply can carry an implicit threat. Such conduct may influence speech without necessarily rising to the level of unconstitutional coercion under existing First Amendment doctrine. 

To get at the heart of the issue, the JAWBONE Act would need to be paired with broader structural reforms. One option would be to narrow the expansive discretionary authority granted to agencies like the FCC and FTC. Vague mandates—such as policing the “public interest” or “unfair or deceptive” practices—create opportunities for regulators to blur the line between legitimate oversight and pressure campaigns directed at speech. 

Congress, and ultimately the Supreme Court, might also reconsider legacy decisions such as Red Lion Broadcasting Co. v. FCC and FCC v. Pacifica Foundation. Those cases afforded broadcasters reduced First Amendment protection based on assumptions about media scarcity that are increasingly difficult to justify in today’s communications environment. Strengthening the constitutional protections available to speech intermediaries would make them less vulnerable to government pressure and reduce opportunities for future jawboning. 

The Cure Is Not the Prevention

The JAWBONE Act reflects a serious attempt to address a real constitutional problem. By allowing citizens to sue government actors directly for monetary damages, it would give those harmed by backdoor censorship a potential avenue for accountability, even when speech platforms decline to challenge government pressure themselves. 

The bill would also help bring government coercion into public view and deter the most serious abuses. But transparency and liability address only part of the problem. To safeguard the marketplace of ideas over the long term, policymakers should also confront the vague, expansive powers that allow federal agencies to pressure speech platforms in the first place. 

Jawboning thrives in the shadows of discretion. Shrink those shadows, and the First Amendment has room to breathe.

The post Government by Raised Eyebrow: The JAWBONE Act and the Problem of Censorship by Proxy appeared first on Truth on the Market.

The European Commission’s new European technological-sovereignty package deserves close attention in Washington, Brussels, and boardrooms on both sides of the Atlantic. Announced June 3, the package aims to strengthen Europe’s capacity in semiconductors, artificial intelligence, cloud computing, open source, and digitized energy infrastructure.

Those are worthy goals. Europe is right to care about resilience, cybersecurity, and whether its firms and public institutions can use advanced digital tools securely. But good objectives do not rescue bad instruments.

For present purposes, the centerpiece is the proposed Cloud and AI Development Act, or CADA. The Commission says CADA would support cloud and AI research, speed data-center deployment, and create a single EU framework to assess cloud and AI “sovereignty.” The ambition is no small thing: Europe wants to at least triple data-center capacity within five to seven years, expand cloud and AI adoption in public and strategic sectors, and reduce reliance on non-EU providers.

The package also includes Chips Act 2.0, an open-source strategy, and an energy-sector digitization roadmap.

These initiatives should be evaluated as proposals, not faits accomplis. CADA and Chips Act 2.0 still must proceed through the EU’s ordinary legislative procedure, with the European Parliament and Council acting as co-legislators. That means amendments, political bargaining, and, ultimately, either a more balanced transatlantic approach or a more deeply protectionist one.

The coming consultation and legislative process therefore matter. This is precisely the stage at which economically grounded criticism can do the most good.

From Security Standards to Nationality Tests

CADA’s most troubling feature is its sovereignty-assurance framework. The Commission’s public description identifies four levels. Level 1 turns on whether data processing and storage occur in the EU. Level 2 requires providers to demonstrate independence from third countries and transparency across the software supply chain. Level 3 requires EU ownership and control, along with additional criteria, such as personnel citizenship—though the Commission may recognize third-country providers. Level 4 requires full transparency and control over the software supply chain, with no third-country interference.

That language may sound administratively tidy. In economic terms, it looks more like a ladder whose upper rungs were built for European firms.

A U.S. cloud provider is not merely being asked to satisfy cybersecurity, reliability, or technical-performance standards. It is being asked to negate the legal consequences of being a U.S. firm. The CLOUD Act requires covered providers to comply with lawful U.S. process for data within their possession, custody, or control, regardless of whether the data are stored inside or outside the United States.

European officials may dislike that feature of U.S. law. They may reasonably treat it as a sovereignty concern. But using it to demote U.S. firms within a procurement or assurance framework is a market-access condition in all but name.

That distinction matters. A neutral security rule asks whether a provider can protect data, maintain resilience, prevent intrusions, document access controls, comply with lawful process, and protect customers from unauthorized disclosure. A nationality-inflected sovereignty rule asks who owns the provider, which legal system may reach it, where its personnel sit, and whether a third country can be said to exert legal influence.

The former measures performance. The latter sorts firms by political identity.

That shift is especially problematic because cloud is a scale market. Hyperscale cloud is not a local utility that lawmakers can duplicate by aspiration. It requires enormous capital investment, energy access, engineering talent, global service reliability, cybersecurity expertise, specialized hardware, developer ecosystems, and continuous innovation.

Steering public- and critical-sector demand away from the most capable providers because they are American would impose an invisible tax on European users. That tax may not appear on the face of the statute. It will show up in higher costs, weaker services, delayed AI deployment, and lower productivity.

The Protectionist Playbook, Updated for the Cloud Era

The Information Technology and Innovation Foundation’s (ITIF) Aegis Project offers a useful framework for understanding the problem. Aegis starts from a simple proposition: America’s leading technology companies are not incidental to U.S. economic power and national security. They are among the principal means by which the United States maintains technological leadership in a world where China pursues state-directed innovation mercantilism.

The project’s Non-Tariff Attack Tracker catalogs policies around the world that are framed as neutral domestic regulation but, in practice, weaken U.S. technology firms, extract value from them, or tilt markets toward national champions.

ITIF describes such measures as anticompetitive market distortions (ACMDs): “government-imposed policies or practices that distort from open trade, competition on the merits, and property rights protections.” Economic research on ACMDs—supported by econometric modeling and pioneered by trade-policy expert Shanker Singham—finds that such distortions misallocate capital, skew risk-taking, reduce productivity and innovation, and ultimately slow economic growth.

CADA’s sovereignty tiers fit this pattern. They are not tariffs. They do not explicitly say that U.S. cloud providers are unwelcome in Europe. In fact, the Commission stresses that much of the market will remain open to international partners.

But competition can be distorted without formally closing the border. If the most valuable public-sector and critical-infrastructure opportunities depend on independence from non-EU legal jurisdictions, EU ownership and control, EU personnel, or the absence of third-country influence, U.S. firms are not competing on the same terms as local firms. They are competing after regulators have imposed a handicap.

ITIF’s analysis of EU cloud-service restrictions identifies similar dynamics in related contexts. Cloud-certification schemes, digital-sovereignty requirements, data-localization mandates, and jurisdictional restrictions can function as de facto barriers to U.S. providers, even when participation remains formally voluntary.

This is a familiar behind-the-border trade problem. It rarely announces itself as protectionism. Instead, it arrives dressed as resilience, autonomy, privacy, procurement integrity, or cybersecurity. If the practical effect is to shift demand from globally competitive U.S. firms to less competitive domestic alternatives, the economic substance is protectionist.

Security Is Not a Passport

The strongest defense of CADA is that Europe has legitimate reasons to worry about legal and geopolitical dependence. That much is true. No serious observer should dismiss concerns about cyber risk, supply-chain fragility, or state access to sensitive data.

The question is not whether Europe may regulate for security. The question is whether the chosen regulatory proxy improves security at a reasonable cost.

On that score, nationality is a crude—and often misleading—proxy. A European-owned provider may have weaker cybersecurity practices than a U.S. provider. A U.S. provider may offer stronger encryption, more transparent compliance processes, better incident response, and more robust auditing than a local competitor.

Likewise, a provider organized under European ownership may still depend on non-European chips, open-source components, cloud tools, developer libraries, or security vendors. Digital infrastructure is not sovereign because a corporate charter is European. It is secure because its architecture, governance, controls, and incentives produce security.

The relevant policy question is one of comparative institutional performance: Which rule addresses the asserted problem at the lowest social cost?

A performance-based security standard targets risk directly. It can require encryption, logging, access controls, incident disclosure, independent audits, redundancy, supply-chain transparency, and enforceable commitments to customers. A nationality- or jurisdiction-based test targets risk indirectly, and often poorly. It excludes or downgrades firms that may be best able to satisfy the underlying security objective while privileging firms that satisfy a political criterion.

The error costs are substantial. In dynamic technology markets, regulations that divert demand from efficient suppliers can weaken investment incentives, slow technology diffusion, and shield incumbents or favored entrants from competitive pressure. That is not a recipe for European competitiveness. It is a recipe for a smaller, more expensive, and less innovative digital ecosystem.

There is also a rule-of-law concern. Public-procurement criteria should be administrable, transparent, and tied to the actual risks that justify them. A sovereignty ladder invites discretionary judgments about control, interference, legal exposure, and the recognition of third-country providers. Those judgments can easily become political bargaining chips.

A firm that satisfies demanding technical standards may still find itself in a lower tier because its parent company is organized under a disfavored jurisdiction. That is not competition policy; it is industrial administration.

Once procurement officials begin treating national origin as a signal of trustworthiness, moreover, that signal rarely remains confined to the most sensitive workloads. Risk categories tend to expand. Exceptional rules have a habit of becoming ordinary practice.

The predictable response from affected firms will be costly corporate engineering. Providers may create local subsidiaries, special governance structures, contractual firewalls, joint ventures, or bespoke sovereign-cloud offerings to satisfy formal requirements. Some of those arrangements may make sense for particular customers. When regulation mandates or privileges them, however, they consume resources that could otherwise go toward product improvements, stronger security, lower prices, or new services.

The cost does not stop with compliance departments. Customers bear it through higher prices, and the broader economy bears it through slower diffusion of better technology.

You Can’t Build an AI Continent on Anti-Cloud Policy

CADA also sits uneasily with Europe’s own AI ambitions. The Commission wants Europe to become an “AI continent.” That aspiration depends on cloud capacity, data-center deployment, advanced semiconductors, developer tools, model-training infrastructure, and a deep bench of complementary services.

For now, U.S. firms lead in many of those areas. One can regret that fact, explain it, or try to narrow the gap. Pretending it away is not an industrial strategy.

A policy that channels strategic demand away from U.S. hyperscalers may appear to create room for European challengers. In the short run, though, it risks raising input costs for the very firms and public agencies that need AI most.

Cloud services are general-purpose inputs. They do not merely affect cloud-provider profits. They affect hospitals, manufacturers, universities, logistics firms, software startups, financial institutions, energy companies, and public administrations. Making those inputs worse or more expensive reduces economy-wide productivity.

Nor is this just a European problem. The United States and Europe face a shared strategic challenge from China’s state-directed technology model. Fragmenting the transatlantic technology stack weakens the allied position. It reduces scale, complicates interoperability, undermines standards cooperation, and diverts political energy from the more important task of disciplining predatory or mercantilist conduct by strategic rivals.

If democratic allies build regulatory walls against each other, China need not build all the walls itself.

Chips Act 2.0 illustrates the same tension. The Commission says the new initiative will strengthen Europe’s semiconductor industry, support design and production, speed permitting, stimulate demand, and deepen strategic partnerships. Many of those goals are sensible.

But semiconductor capacity is not built in isolation from U.S. firms. American companies design, develop, supply, and invest across the European semiconductor ecosystem. If EU technology sovereignty becomes a preference for European ownership rather than an effort to build resilient allied capacity, it will undermine the investment and cooperation Europe needs.

Competition Policy or Competitor Policy?

There is a deeper concern. Current market shares and existing dependencies can tempt regulators to treat successful foreign firms as a problem to be managed rather than as engines of competition and innovation.

That temptation has surfaced repeatedly in European digital policy. The Digital Markets Act and related initiatives often begin with the premise that large U.S. technology firms are too powerful and then translate that judgment into regulatory obligations governing product design, data use, interoperability, contracting, and business models.

To be sure, there are signs that EU courts are placing greater emphasis on evidence, proportionality, dynamic competition, and rigorous legal reasoning when reviewing European Commission antitrust and digital-regulation decisions. CADA should be judged by those same standards.

Dynamic competition asks a different question from static market-share analysis: Which rules are most likely to maximize innovation, investment, entry, and productivity over time?

That framework does not assume that today’s largest suppliers deserve permanent market positions. But neither does it assume that regulators can improve outcomes by steering demand toward politically preferred suppliers. The goal should be competition on the merits.

If a European cloud provider offers better security, lower costs, stronger service, or superior compliance, it should win customers. If a U.S. provider offers those advantages, it should win customers. Public procurement should not become a vehicle for industrial favoritism masquerading as sovereignty.

The risk is not merely that U.S. firms lose business. The greater risk is that European users lose access to best-in-class tools and that European innovators must build on a less capable technological foundation.

The costs of that choice compound over time. Lower-quality cloud services slow AI adoption. Slower AI adoption weakens productivity growth. Weaker productivity growth reduces the fiscal and political capacity needed to invest in genuine resilience.

Protectionism marketed as sovereignty can leave Europe less sovereign in the ways that matter most.

Sovereignty Without Protectionism

There is a constructive alternative. Europe can pursue resilience without discrimination. It can define security requirements in technological and contractual terms rather than nationality-based ones.

For example, it can require providers to disclose supply-chain dependencies, submit to independent audits, maintain EU-based redundancy, support customer-controlled encryption, report government-access requests where lawful, and provide robust exit and portability tools. It can reserve the most sensitive national-security workloads for narrowly defined sovereign environments without turning broad public-sector procurement into a buy-European program.

For its part, the United States should take European concerns seriously while making clear that discriminatory market-access conditions are unacceptable. Diplomacy should come first. The goal is not to turn every regulatory disagreement into a trade war.

At the same time, as ITIF’s recent Section 301 report argues in the broader EU context, the United States should be prepared to identify, document, and respond to discriminatory policies that burden U.S. technology firms. Serious negotiations are more likely when both sides understand that behind-the-border discrimination carries consequences.

A transatlantic bargain should be possible. The United States and Europe could work toward mutual assurances on lawful access, transparency surrounding government requests, cybersecurity certification, procurement neutrality, cloud portability, and trusted-vendor criteria that distinguish allies from strategic adversaries.

They could also coordinate semiconductor incentives and AI-infrastructure investments so that public subsidies reinforce allied supply chains rather than fragment them. That approach would strengthen resilience without sacrificing competition.

The key is to avoid turning legitimate sovereignty concerns into a general license for industrial policy. Europe’s problem is not that American cloud firms are too capable. Europe’s problem is that too many of its own firms, agencies, and institutions have been slow to adopt digital technologies effectively.

Punishing the suppliers most capable of helping close that gap is an odd way to solve it.

Don’t Mistake Sovereignty for Strength

The Commission’s technology package remains only a proposal. That is good news. It means policymakers still have time to separate genuine resilience measures from protectionist market design.

CADA’s sovereignty-assurance framework, as currently described, raises serious concerns because it risks turning legal jurisdiction and corporate ownership into decisive competitive criteria. That would burden U.S. providers, reduce European access to leading cloud and AI infrastructure, and weaken the transatlantic technology ecosystem at precisely the wrong moment.

Europe should build. It should invest. It should streamline permitting, expand energy capacity, support research, and accelerate technology adoption. It should also protect sensitive data and critical infrastructure.

But it should pursue those goals through performance-based, evidence-based, and nondiscriminatory rules. Strategic autonomy should mean the ability to choose among excellent technologies from trusted allies, not an obligation to prefer a less capable domestic alternative because it carries the right passport.

If the European Union wants to become an AI continent, it should resist the temptation to become a regulatory island.

The path to technological sovereignty runs through competitiveness, not protectionism.

The post Europe’s Cloudy Judgment appeared first on Truth on the Market.

In March, the International Center for Law & Economics (ICLE) submitted comments to the European Commission on the General Data Protection Regulation (GDPR) and ePrivacy provisions of the Digital Omnibus. The GDPR is the European Union’s main privacy law. The ePrivacy rules govern confidentiality of communications and device-access rules, including the cookie-consent regime that has trained a continent to click “accept” while learning nothing substantial. 

We made four core arguments.

First, policymakers should adopt the proposed entity-relative clarification of the personal-data definition—that is, they should ask whether the particular organization holding the data can realistically identify someone, not whether someone, somewhere, with some imagined tool, might be able to do so.

Second, the package’s artificial-intelligence (AI) provisions represented a necessary legislative settlement of questions that the European Data Protection Board (EDPB), the EU body that coordinates national privacy regulators, had deliberately left unresolved.

Third, the proposed cookie-consent reforms pointed in the right direction, but did not go far enough.

Finally, the package’s greatest weakness was its silence on enforcement architecture. Without institutional reform, we argued, the same authorities that had interpreted the GDPR into a “law of everything” would read the new exemptions just as narrowly. 

Three months later, the two sister components of the Digital Omnibus have diverged sharply. The AI Omnibus (COM(2025) 836) reached a provisional trilogue agreement on May 7. A trilogue is the closed-door negotiation among the Commission, Parliament, and Council that often determines the final shape of EU legislation. The Data Omnibus (COM(2025) 837)—which contains the GDPR and ePrivacy reforms—remains before the Council, where the Cypriot Presidency has circulated successive compromise texts, most recently on May 21 (Council document 9547/26). 

That latest compromise deletes three of the Commission’s four principal GDPR reforms: the entity-relative personal-data test, the relocation of cookie consent into the GDPR, and the legitimate-interest basis for AI processing. It also removes the Commission’s proposed authority to define pseudonymization criteria. Pseudonymization means replacing direct identifiers, such as names, with substitutes, while keeping the possibility of re-identification under controlled conditions. 

What remains of those reforms has largely migrated into recitals and EDPB guidance. Recitals are the explanatory passages that accompany EU laws. They can influence interpretation, but they are not the same as binding operative text. At the same time, the compromise expands the EDPB’s own mandate. 

The concern at the heart of our March comments—that textual reform without enforcement reform would underperform—now looks almost understated. The Council appears poised to deliver less reform than the Commission proposed, while leaving the EDPB stronger than before. 

This post examines how the Council compromise affects each of the issues addressed in our comments, and what it means as the file moves to the European Parliament. 

The Referee Takes the Field

Our comments urged adoption of the amended Article 4(1), which would have codified the entity-relative approach to identifiability reflected in EDPS v. SRB and Breyer. We also supported Article 41a, which would have authorized the Commission—after consulting the EDPB—to specify, by implementing act, when pseudonymization places data outside the scope of the GDPR. An implementing act is a legal instrument through which the Commission sets technical or practical rules under authority granted by legislation. 

That allocation of authority mattered. We warned against leaving the scope of the law to be defined solely by the body that enforces it. 

The Council deleted both provisions. 

The entity-relative test disappeared from operative Article 4(1). Its substance survives only in Recital 27a, which states that identifiability “should be assessed by the controller or the processor, considering the actual technical, organisational and legal capabilities” of that controller or processor. In GDPR terms, a controller decides why and how personal data is processed; a processor handles data on the controller’s behalf. 

The reference to “actual … capabilities” is welcome. It addresses the recurring tendency to treat purely hypothetical means of identification as legally relevant. But the clarification now appears only in a recital, rather than in binding legal text. 

Article 41a fared even worse. The Council eliminated the Commission’s implementing-act authority entirely. Renumbered as Article 29a, the provision now directs the EDPB to issue an opinion on pseudonymization and anonymization within 12 months of the regulation’s entry into force. 

Peter Craddock greeted the new Article 29a with a pointed question: “Possibly the most useless legal provision ever proposed?” On our reading, it is worse than useless. A provision originally intended to check the EDPB’s steadily expanding jurisdiction—by giving the Commission authority to define the contours of pseudonymization—now does the opposite. It more explicitly empowers the Board to define the limits of its own jurisdiction. The revised Article 70(1) confirms the point by assigning the Article 29a opinion to the EDPB. 

That institutional shift matters because the Commission and the EDPB have not always agreed on how broadly data-protection rules should reach. In EDPS v. SRB, the Commission intervened to push back against a maximalist interpretation of identifiability. Under the Council text, by contrast, the EDPB will determine what qualifies as effective pseudonymization. 

This is the same body that interpreted Article 5(3) of the ePrivacy Directive so broadly that URL parameters and ordinary browser transmissions became consent-triggering forms of “access.” Expecting a notably pragmatic approach to pseudonymization would be optimistic. 

Of the two deletions, the loss of Article 41a matters more in practice. The Article 4(1) amendment was symbolically important, and we continue to believe it should be restored. Standing alone, though, it might not have changed an enforcement culture resistant to the idea that “personal data” has meaningful limits. 

Article 41a addressed the institutional problem directly by vesting interpretive authority in a body other than the EDPB. That is precisely why the EDPB and the European Data Protection Supervisor (EDPS), the EU’s independent data-protection authority for EU institutions, called for its deletion in their February joint opinion. It is unfortunate that the member states chose to follow that advice. 

The Deal That Unmade Itself

We recommended adopting both Article 88c, which would have created a legitimate-interest basis for AI development and operation, and Article 9(2)(k), which would have created a derogation for special-category data that appears incidentally in AI training. “Legitimate interest” is one of the GDPR’s lawful bases for processing personal data. Special-category data includes especially sensitive information, such as data revealing health, race, ethnicity, political opinions, religious beliefs, biometric identifiers, or sexual orientation. 

Article 88c has disappeared from the operative text, but much of its substance survives in Recital 33a. Most of the Commission’s language remains, albeit with important changes. Processing that the Commission said “may be pursued for” legitimate interests now merely “may be regarded as” carried out for a legitimate interest. The unconditional right to object—the central safeguard, which we described in March as stronger than the ordinary Article 21(1) standard—has been deleted, along with the enhanced-transparency requirement. The Council also added a caveat stating that the recital “does not affect the obligation … to choose the most appropriate lawful ground.” 

The result is a return to the ordinary Article 6(1)(f) balancing test, with EDPB Opinion 28/2024 serving as the de facto framework. That opinion was notable for what it did not resolve, leaving broad discretion to enforcement authorities. Both sides of the Commission’s original bargain have therefore disappeared: the clearer legal basis that AI developers were supposed to receive, and the unconditional objection right that data subjects—the people whose data is processed—were supposed to receive in return. 

To be fair, the distinction between operative text and a recital may matter less in practice than the attitude of the authorities enforcing it. A supervisory authority determined to reach a particular result can often interpret around either one. But that observation supports, rather than undermines, the point we made in our comments: institutional incentives matter at least as much as statutory language. 

Article 9(2)(k) survived, albeit in narrowed form.

The Council added the qualifier “incidental and residual,” and moved the accompanying conditions into a new Article 9(5). Controllers must take measures to avoid collecting special-category data; erase such data once identified; and, where erasure is impossible or manifestly disproportionate—for example, because the information has been memorized by a model—protect it against further processing, inference, or disclosure. They must also document the process. The provision expressly excludes data collected through prompts during deployment. 

Even in this narrower form, Article 9(2)(k) remains useful. Web-scale AI training makes the incidental collection of sensitive data effectively unavoidable. The provision answers a question that legislators can no longer sidestep. Indeed, it may be the only significant AI-related GDPR reform still standing in the package. 

The AI Omnibus, meanwhile, is essentially finished. 

The agreed text extends the bias-detection carveout—formerly Article 10(5) of the AI Act, and now a standalone Article 4a—beyond providers of high-risk systems to deployers and to non-high-risk AI systems and models. In AI Act terms, providers develop or place AI systems on the market, while deployers use them. The carveout preserves the “strictly necessary” threshold and the existing safeguards. It operates through Article 9(2)(g) GDPR; the AI Act supplies the required basis in EU law, while Article 4a supplies the accompanying safeguards. 

The agreement also postpones the compliance dates for high-risk systems, moving Annex III obligations to Dec. 2, 2027, and product-embedded AI obligations to Aug. 2, 2028. At the same time, it accelerates the Article 50 watermarking requirements to Dec. 2 of this year and adds a new Article 5 prohibition on AI-generated child sexual-abuse material and non-consensual intimate imagery. 

Taken together, the two files produce a striking asymmetry. The deliberate use of sensitive data for bias detection—the relatively uncommon case—now rests on firm legislative ground. The incidental presence of sensitive data in training datasets—the ubiquitous case—remains trapped in a contested file, within a provision the Council has already narrowed. 

Meet the New Cookie Rules, Same as the Old Cookie Rules

On ePrivacy, we were blunt in March. The Commission’s proposal was the weakest part of the package. It created a two-track regime, moving personal-data-related device access into a new Article 88a of the GDPR while leaving non-personal data under Article 5(3) of the ePrivacy Directive. Its analytics exemption was too narrow to cover standard third-party analytics services, and it omitted several low-risk exemptions that should have been obvious candidates, including fraud prevention, advertising measurement, frequency capping, and contextual advertising. 

The Council solved the two-regime problem by abandoning the reform altogether. 

Article 88a is gone. Device access remains governed by Article 5(3) of the ePrivacy Directive, albeit in revised form. That means seven years of expansive EDPB interpretations of Article 5(3)—covering everything from cookies and tracking pixels to URL parameters—remain the starting point. Whatever the revised exemptions ultimately say, the EDPB will continue to play a decisive role in determining how broadly or narrowly they are interpreted. 

The revised exemptions move toward some of the reforms we recommended, only to hedge them so heavily that they risk becoming ineffective in practice. 

The audience-measurement exemption now permits analytics conducted “by a third party acting on the provider’s behalf.” That directly addresses the analytics problem we identified and could help resolve the long-running Google Analytics disputes. But the exemption applies only if the data remains anonymous and aggregated, and is neither combined with other data nor shared. 

Fraud prevention now appears within the security exemption, but the drafting limits it to protecting the “security of the interface.” As Craddock observed, fraud prevention typically aims to protect the broader service from fraudulent users, rather than the interface itself. Under the current text, “in most cases, fraudsters will then need to give their consent before anything can be done against them.”  

The contextual-advertising exemption fared even worse. It appeared in the February compromise text but disappeared by May, even though the joint opinion of the EDPB and EDPS had itself recommended exempting contextual advertising. Contextual advertising targets ads based on the content a user is viewing, rather than by tracking that user across sites and services. 

The browser-signals provision survives. Renumbered as Article 88a, it now applies to operating-system providers as well as browsers. The provision also includes a one-click refusal mechanism, a six-month cooling-off period following refusal, and a 24-month transition period. National GDPR supervisory authorities would serve as ePrivacy enforcers. 

The comparison we drew with the United Kingdom has, if anything, become sharper. 

The cookie provisions of the Data (Use and Access) Act 2025 took effect Feb. 5. Rather than relying on a single “strictly necessary” exemption, the law creates five exemptions, including an analytics exemption that expressly permits third-party providers acting as processors under an opt-out model. It also authorizes the secretary of state to create additional exemptions through secondary legislation.  

Last month, the Information Commissioner’s Office (ICO), the United Kingdom’s data-protection regulator, went further, formally advising the government to remove low-risk advertising activities from consent requirements altogether. The ICO proposed a “first-party framework” covering ad delivery, measurement, frequency capping, brand safety, and ad-fraud prevention, while retaining consent requirements for cross-service tracking and profiling. The proposal was evidence-based, supported by cost analysis, and tested through citizen juries. 

The ICO now also operates under a statutory duty to balance privacy concerns against innovation, competition, and crime prevention. We do not endorse every conclusion the ICO reached. One could argue that the cleaner solution is simply to repeal the cookie-consent rule and allow the GDPR’s risk-based framework to govern device access. Even so, the contrast between the British and European approaches is stark. 

Our recommendation therefore remains unchanged. The best solution is to repeal Article 5(3) of the ePrivacy Directive and allow the GDPR to govern device access without a parallel consent regime. Short of that, the European Parliament should at least detach the fraud-prevention exemption from the concept of “interface” security, remove the restrictions that make the analytics exemption largely unusable, and restore the contextual-advertising exemption. 

The Changes You Might Have Missed

The compromise also reshaped three quieter corners of the file. Two concern provisions that our March comments did not address in detail. All three, however, follow a familiar pattern: legislative reforms survive, but key questions about their scope are increasingly delegated to the EDPB. 

The EDPB Defines Science

The compromise retains the new GDPR definition of “scientific research” in Article 4(38), but in a form largely rewritten to the EDPB’s specifications. Research must now be “conducted in an autonomous and independent manner,” follow recognized methodological standards, and produce “verifiable and transparent results.” 

The Commission’s original language acknowledging that research may support innovation and commercial objectives has been removed from the operative text and relocated to Recital 28. At the same time, the EDPB has published Guidelines 1/2026 on scientific research, currently open for public consultation, which seek to define the concept through guidance.

The result is familiar. The legislation preserves a broader category, but much of the practical work of deciding who qualifies for it will occur through EDPB interpretation.

Yes, But Only If You Don’t Use It

The Commission proposed, and the Council retained, a new Article 9(2)(l) allowing biometric verification when both the biometric data and the means of verification remain under the data subject’s sole control. Biometric verification confirms that a person is who they claim to be by comparing a live biometric sample—such as a face or fingerprint—to a stored template.

The Council added one further limitation: the verification must be “one-to-one,” meaning the system compares a user against a single stored biometric template rather than searching a larger database.

The more significant constraint may lie in Recital 34. As revised by the Council, the recital instructs controllers to prioritize authentication methods that do not use biometric data and to select the “less intrusive” option whenever two methods are equally effective.

That language comes directly from EDPB Opinion 11/2024 on facial recognition at airports. Notably, the opinion rejected biometric verification even where travelers had a genuine choice among authentication methods. The operative rule therefore creates a new permission, while the accompanying recital imports reasoning that could substantially limit its practical use. 

The Algorithm Still Says No

The Council retained the Commission’s broader rewrite of Article 22 on automated decision-making, but reverted to the familiar “right not to be subject” formulation. In plain English, Article 22 governs when consequential decisions—such as some credit, employment, or eligibility decisions—may be made solely by automated systems without meaningful human involvement.

By returning to the “right not to be subject” language, the Council preserved the prohibition-in-principle approach endorsed by the Court of Justice of the European Union (CJEU) in SCHUFA. 

The Commission’s most useful clarification did not survive in operative text. Its proposal would have made clear that a decision can be “necessary” for the performance of a contract even when a human could theoretically have made the same decision. That clarification now appears only in Recital 38. 

Meanwhile, the revised Article 70(1) assigns the EDPB responsibility for specifying the criteria governing Article 22 profiling decisions. 

Once again, the pattern is hard to miss. The legislation settles some questions, moves others into recitals, and leaves the EDPB with a larger role in determining how the final framework will operate in practice. 

The Board Wins

In our March comments, we argued that the Digital Omnibus was fundamentally incomplete. The same authorities whose interpretations helped create the current dysfunction would remain responsible for interpreting whatever reforms the legislature enacted. The package contained no mechanism to discipline that process.

The Council compromise has done more than validate that concern. It has expanded the EDPB’s formal authority in the very same text that removes many of the Commission’s proposed legislative settlements.

Recital 40a instructs national supervisory authorities to ensure that national guidance on matters covered by EDPB guidelines “does not contradict” the Board’s positions. Nothing comparable appeared in the Commission’s proposal. The result is a new form of vertical lock-in, giving EDPB interpretations greater influence over national enforcement.

The revised Article 70(1) likewise expands the Board’s role. The EDPB is now tasked with developing guidelines that specify the criteria for Article 22 profiling decisions and Article 32 security measures. It will issue the Article 29a opinion on pseudonymization and anonymization. It will also gain authority to establish data-protection-impact-assessment (DPIA) and breach-notification lists directly, rather than merely proposing them to the Commission. A DPIA is a risk assessment that organizations must sometimes conduct before engaging in higher-risk data processing.

Viewed together, these changes raise an obvious question: where do the substantive issues that the Commission originally sought to settle now reside? 

In other words, many of the questions that the Commission attempted to answer through legislation have been returned to the EDPB, either directly or indirectly.

Accountability remains limited.

In WhatsApp Ireland v. EDPB, the Grand Chamber confirmed in February that the Board’s binding decisions under Article 65 may be challenged before the EU courts. That is a genuine improvement. But it applies only to the relatively rare instances in which the EDPB issues formally binding decisions. 

The situation is different for the guidelines and opinions that do most of the practical work of shaping enforcement. In Meta v. EDPB, the General Court dismissed a challenge to the Board’s consent-or-pay opinion because an EDPB opinion “does not produce binding legal effects.” The appeal remains pending.

That distinction matters because the Omnibus’ substantive compromises have increasingly migrated into precisely those instruments—guidelines, opinions, and interpretive documents—that courts generally will not review.

That development makes every element of the structural reform agenda outlined in our March comments more urgent. Mandatory proportionality assessments for EDPB outputs, an independent multidisciplinary review mechanism, and a clearer standard for judicial review all matter more after the Council’s compromise than they did before.

Parliament’s Last Chance to Legislate

The Council may reach a general approach this month, before Cyprus hands the presidency to Ireland on July 1. In the European Parliament, responsibility is shared between the Committee on Industry, Research and Energy, led by rapporteur Aura Salla of the European People’s Party, and the Committee on Civil Liberties, Justice and Home Affairs, led by rapporteur Marina Kaljurand of the Socialists and Democrats. 

Kaljurand has been openly skeptical. At the committee’s January hearing, she warned that allowing AI systems to process sensitive data for bias detection would mean the “technological neutrality of the GDPR will no longer exist.” The Socialists and Democrats group has likewise rejected what it calls “unacceptable deregulation and weakening of the EU’s digital rules,” while groups to the right of the European People’s Party have pressed for the package to go further. As of this writing, the committees had yet to settle their working arrangements, and adoption realistically appears likely to slip into 2027. 

The open question is whether Parliament will restore any of the Commission’s reforms to binding text.

For Parliament—and for member states still negotiating the general approach—our recommendations follow directly from the analysis above:

• Restore the entity-relative test to operative Article 4(1). Recital 27a’s “actual technical, organisational and legal capabilities” language belongs in binding text. 
• Restore the Commission’s implementing-act authority over pseudonymization criteria. Article 29a, in its current form, should be deleted. It asks the body whose jurisdiction is at issue to define the limits of that jurisdiction. 
• Restore Article 88c as an operative legal basis for AI development and deployment, together with the unconditional right to object that the Council removed. Retain Article 9(2)(k), even in its narrowed “incidental and residual” form. 
• Keep Article 9(2)(l), but rewrite Recital 34. A permission in operative text should not be paired with a recital instructing authorities to disfavor the very processing the permission allows. 
• On ePrivacy, repeal Article 5(3) without replacement and let the GDPR govern device access. Short of that, detach fraud prevention from the “security of the interface,” remove the restrictions that make the analytics exemption impractical, and restore the contextual-advertising exemption. 
• Delete Recital 40a and reverse the Article 70(1) mandate expansion. A reform prompted by the costs of unaccountable interpretive power should not end by enlarging it. 
• Take up enforcement reform. The structural agenda of our March comments—separating investigation from adjudication, creating multidisciplinary review of consequential decisions, requiring mandatory balancing of all affected rights and interests, and enabling review of EDPB quasi-legislation—is still missing from the Digital Omnibus. 

The Digital Omnibus began as a promise to make EU data law more proportionate, predictable, and compatible with European competitiveness. Six months in, the Council’s answer is a text that keeps the procedural trimmings, moves the substance into recitals and EDPB opinions, and strengthens the very institution whose interpretive practice made reform necessary in the first place.

Parliament now has a chance to prove that EU data-protection law is still made by legislators, not outsourced to the referee.

The post EU Digital Omnibus Hands the Wheel to the Referee appeared first on Truth on the Market.

The underlying facts are not especially favorable to Perplexity. In granting a preliminary injunction, Judge Maxine Chesney of the U.S. District Court for the Northern District of California found “strong evidence” that Perplexity violated both the federal CFAA and California’s analogous statute. According to the court, Perplexity continued accessing Amazon’s systems after receiving a cease-and-desist letter and deliberately evaded the technical measures Amazon deployed to block that access. The 9th Circuit stayed the injunction pending appeal. 

The doctrinal question is the easy one. Amazon will probably win, and probably should. It is also the less interesting question. 

The harder and more consequential issue is whether the CFAA is the right body of law to govern this kind of dispute at all. More broadly, it raises a recurring problem in technology law: whether it is sustainable to keep asking statutes written for the technological realities of the mid- and late-20th century to govern technologies their authors could not have anticipated. 

We think the answer to both questions is no. As Greg Dickinson puts it in his masterful article, “Law Proofing the Future”:

Technological breakthroughs provoke wonder, then fear, then legislation. The resulting legal regimes entrench incumbents, suppress experimentation, and displace long-standing legal principles with bespoke but brittle rules. . . . [Meanwhile,] the most powerful tools for governing technological change—the general-purpose tools of the common law—are in fact already on the books, long predating the technologies they are now called upon to govern, and ready also for whatever the future holds in store.

The interests Amazon seeks to protect are real. But they are fundamentally interests in property and contract, and courts developed the core principles governing those interests long before Congress enacted the CFAA. When statutes track those common-law principles, they often work well. When they depart from them—or prevent the sort of incremental adaptation that characterizes the common law—they tend to generate exactly the kind of doctrinal strain the CFAA now exhibits. 

The lesson for agentic AI is not that Congress needs to enact a new statute. The legal system already possesses a framework capable of absorbing these new facts. Under current political and institutional conditions, any new legislation is more likely to depart from that framework than to reinforce it. 

When a Contract Dispute Wears a CFAA Costume

The factual record is unflattering enough to Perplexity that, under existing doctrine, Amazon has a strong CFAA case. According to the complaint, Perplexity’s Comet browser transmits the same user-agent string as Google Chrome, making its automated activity inside an authenticated Amazon session difficult to distinguish from that of a human shopper. When Amazon deployed a technical block targeting Comet in August 2025, Perplexity allegedly pushed a software update within 24 hours that changed Comet’s fingerprint and restored access. Amazon sent a cease-and-desist letter to Perplexity’s CEO on Oct. 31, 2025. Perplexity declined to comply and instead published a blog post titled “Bullying is not innovation.” Amazon sued four days later. 

The doctrinal question is whether an authenticated session—but one conducted contrary to terms of service that prohibit third-party automation and after Amazon revoked access by contractual and technical means—is the kind of environment in which an automated commercial agent acts “without authorization” under the CFAA. After the Supreme Court’s 2021 decision in Van Buren v. United States, authorization is, roughly, a function of whether the accessed system is “gates-up” or “gates-down.” Here, as Judge Chesney put it, Comet accessed Amazon “with the Amazon user’s permission, but without authorization by Amazon.” 

The Supreme Court in Van Buren took pains to read the statute narrowly, partly because a broader reading could turn routine terms-of-service violations into federal computer crimes. In particular, after Van Buren, it is hard to argue that a user’s authorized session becomes a federal computer crime simply because the user also does something that violates the terms of service. Amazon’s strongest theory is therefore not the bare terms-of-service violation, which, after Van Buren and the 9th Circuit’s decision in hiQ II, might not be enough on its own. 

Instead, Amazon’s case rests on three facts: Amazon deployed a technical countermeasure specifically targeting Comet at the point of access; Perplexity pushed an update within 24 hours to defeat it; and Comet continued to spoof a Chrome user-agent string to evade detection. On that record, Amazon’s claim is not merely that Perplexity breached a contract. It is that Amazon closed a gate, and Perplexity deliberately broke through it. That is much closer to the paradigmatic CFAA case preserved by Van Buren than to the terms-of-service cases the Supreme Court has narrowed. 

So Amazon probably wins.

The problem is that Amazon’s path to victory requires the CFAA to do work that exposes its awkward conceptual fit. The “gate” Amazon seeks to protect under the CFAA is one it erected because its underlying contractual prohibition on agentic access was, by itself, inadequate. Detection was costly, circumvention was fast, and ordinary contract remedies were too slow to matter. Pressing the CFAA into service this way is understandable, but it conflates the intrusion claim the statute was designed to police with the property-and-contract claim actually at issue. 

That raises the larger question: Why is this dispute being resolved under a federal anti-intrusion statute drafted at the dawn of personal computing, rather than through the bodies of law—property, contract, trespass, and duty of care—that courts normally use to allocate responsibility for such harms? And what should the answer tell us about how the law should govern agentic AI? 

The Matthew Broderick Theory of Legislation

The CFAA traces, somewhat embarrassingly, to a Hollywood blockbuster. 

WarGames premiered in June 1983. Within days, President Ronald Reagan reportedly raised the film’s hacking scenario with the Joint Chiefs of Staff. Congressional hearings followed within a year, and Congress enacted the statute that became the CFAA in 1986. The threat model was straightforward: an outsider breaking into a discrete government or defense computer system. The statute’s core concepts—“access without authorization,” “exceeds authorized access,” and “protected computer”—were drafted with that scenario in mind. 

Yet Congress was not entering a legal vacuum. By 1986, the common law had spent centuries developing doctrines for analogous disputes. Trespass to chattels and conversion addressed unauthorized use of another’s property. Nuisance governed conduct that interfered with the productive use of resources. Principal-agent doctrines distinguished between actions that bound a principal and actions that exceeded an agent’s authority. 

Indeed, before the CFAA became the dominant framework, courts often analyzed unauthorized automated access through the common-law doctrine of trespass to chattels. Cases such as CompuServe v. Cyber Promotions (1997) and eBay v. Bidder’s Edge (2000) reached results that look remarkably similar to what Amazon seeks here. The difference is that they did so through doctrines grounded in property and contract. The common law already had tools for addressing these disputes. In important respects, the CFAA displaced those tools rather than allowing them to continue evolving. 

The result is a textbook case of technological panic producing bad legislation. As Kevin Frazier recently observed (discussing Dickinson’s “Law Proofing the Future”): 

Enacted in 1986 to address hacking of government systems, the CFAA soon expanded into a sweeping prohibition on ‘unauthorized access’ to any computer connected to the internet—which, in practice, meant nearly everything. . . . What began as a targeted fix for a Cold War fear became a vague, overbroad, and ill-fitting statute that chilled ordinary activity. It is precisely the kind of misfire we should expect when lawmakers legislate in haste against technologies they barely understand.

The dispute in Amazon v. Perplexity bears little resemblance to the intrusion Congress had in mind in 1986. No one is breaking into a military computer or bypassing a security perimeter to obtain forbidden information. Instead, a commercial software agent is acting on behalf of a paying customer and carrying out transactions the customer is entitled to perform on a platform that invites the customer’s business. 

Amazon’s interests are real. It has legitimate concerns about advertising-impression integrity, fraud detection, authenticated-session security, and the economics of its affiliate program. But those are not the interests the CFAA was designed to protect. They are interests that fit far more naturally within contract, property, and tort law.

You Can’t Give Away Rights You Don’t Have

Absent the CFAA mismatch, the Amazon-Perplexity dispute seems relatively straightforward. 

Amazon owns the servers. Authenticated customer accounts run on those servers. The merchant, affiliate, and advertising relationships layered on top of the platform are contractual. The right to determine the terms of access—including whether and on what terms automated agents may operate within authenticated sessions—is part of the bundle of rights Amazon possesses and may choose to allocate. 

None of that requires the CFAA. Trespass to chattels, breach of contract, tortious interference with business relationships, and unfair-competition law all potentially reach the conduct Amazon challenges. Amazon’s reliance on the CFAA instead reflects the statute’s procedural and strategic advantages—federal jurisdiction, a criminal-law backstop, and enhanced remedies—not the nature of the underlying interests at stake. 

The strongest defense of agentic access rests on user consent. Perplexity’s argument, at bottom, is that its AI agent is simply doing what the end user has authorized it to do. If a user may browse Amazon, why can’t the user’s software agent browse Amazon on the user’s behalf? 

The common law has long had an answer. A guest’s permission to enter property does not automatically entitle the guest to bring along commercial agents whom the property owner has not authorized. Justice Clarence Thomas captured the principle succinctly in his dissent in Van Buren: 

As the Second Restatement of Torts explains, “[a] conditional or restricted consent to enter land creates a privilege to do so only in so far as the condition or restriction is complied with.” . . .

. . . What is true for land is also true in the computer context; if a company grants permission to an employee to use a computer for a specific purpose, the employee has no authority to use it for other purposes.

Use beyond the scope of consent renders the visitor a trespasser. 

The same logic appears in the common law of agency. An agent cannot possess greater authority than its principal. The principle traces to the ancient maxim nemo dat quod non habet: one cannot give what one does not have. Amazon’s contractual prohibition on third-party agentic access is therefore the relevant constraint. A user’s decision to employ Perplexity’s agent cannot expand the underlying property and contractual rights the user received from Amazon in the first place. 

The 9th Circuit has already confronted—and rejected—essentially the same user-consent argument. In Facebook v. Power Ventures, the court held that a third-party aggregator’s continued access to Facebook on behalf of users who voluntarily supplied their login credentials was nevertheless “without authorization” under the CFAA once Facebook issued a cease-and-desist letter and the aggregator circumvented Facebook’s technical restrictions: 

Similarly, for Power to continue its campaign using Facebook’s computers, it needed authorization both from individual Facebook users (who controlled their data and personal pages) and from Facebook (which stored this data on its physical servers). Permission from the users alone was not sufficient to constitute authorization after Facebook issued the cease and desist letter.

Once platform-level permission has been revoked, the court explained, neither technological workarounds nor user authorization can restore it. The implication here is straightforward: authorization from the user is not a substitute for authorization from the platform. A third party that acquires a user’s credentials acquires no greater rights against the platform than the user possesses. 

It’s Not ‘Bullying’ to Protect One’s Property

Perplexity’s public response illustrates a broader problem with how parts of the AI industry frame these disputes. After receiving Amazon’s cease-and-desist letter, Perplexity published its “Bullying is not innovation” blog post. The substance of the argument is that legal constraints on Perplexity’s preferred mode of operation are themselves illegitimate—that requiring an AI agent to identify itself, comply with contractual restrictions, or respect technical barriers constitutes “bullying.” 

That framing gets things backward. Describing the enforcement of property and contract rights as bullying is rhetorical sleight of hand. It asks the legal system to weaken the very institutions that make commercial cooperation and innovation possible. If anything, the stronger claim runs in the opposite direction: demanding unauthorized access to someone else’s property is not “innovation.” 

The concept of “permissionless innovation,” when applied accurately, denotes autonomy from government mandates, not an exemption from private contractual obligations. There is no contradiction in arguing that AI agents should operate without regulatory pre-approval while maintaining that they must obtain consent from the platforms they engage with. Far from being a substitute, a robust system of private ordering serves as the essential foundation for innovation that is truly permissionless.

Indeed, complying with basic conditions of access should be table stakes. Among other things, agentic browsers introduce attack surfaces that did not previously exist, including indirect prompt injection and session-hijacking risks of the sort that have come to be called “CometJacking”—literally named after Perplexity’s browser. Platforms have a legitimate interest in knowing which sessions are being driven by software agents, what those agents are doing, and how to respond when something goes wrong. None of this is rendered moot because it is new technology doing the accessing.

The Supreme Court’s decision in American Broadcasting Cos. v. Aereo offers a useful analogy. Aereo built a technically ingenious system consisting of thousands of tiny antennas, each assigned to an individual user. The architecture was designed to allow Aereo to retransmit broadcast television without paying the licensing fees that a conventional retransmitter would have owed to the broadcaster. 

The company’s argument was essentially that its novel technical design changed the legal character of the underlying conduct. The Supreme Court disagreed. 

The broader lesson of Aereo should apply here. A defendant cannot avoid legal obligations merely by inserting a clever technical architecture between itself and the conduct at issue. Agentic AI is, in this sense, the new Aereo: a novel technical layer placed on top of conduct that would otherwise require the platform’s consent. The novelty of the architecture should not alter the legal analysis. 

The Statute Trap

The CFAA is not unique in routing around the common law. Modern technology law has repeatedly taken problems the common law could have absorbed and recast them into statutes tailored to the anxieties, assumptions, and political pressures of a particular technological moment. The results vary. Some statutes ossify into broad immunity. Others stretch far beyond their original purpose. A few work tolerably well because they preserve enough flexibility for courts to adapt them over time. 

Agentic AI is now provoking the same legislative instinct. But before Congress reaches for the drafting pen, it is worth examining how similar efforts have fared. 

Section 230: The Immunity Machine

The CFAA’s drift away from the common law has a structural counterpart in another statute whose pathologies have come to dominate digital-policy debates: Section 230 of the Communications Decency Act. 

The common law of intermediary liability had already developed coherent answers to many of the problems Section 230 purported to solve. Courts had long grappled with analogous questions through doctrines governing innkeeper liability, premises liability, dram-shop liability, and the duty-to-control doctrine. Across these doctrines, a common principle emerged: liability generally turns on whether the intermediary is the least-cost avoider of the harm, whether the harm was foreseeable, and whether reasonable steps could have mitigated it. 

These are not abstract concepts. They are centuries-old institutions for allocating responsibility among parties whose conduct predictably affects one another. As we have noted elsewhere: 

While it was once in fashion to proclaim the Internet a wholly unique invention to which traditional laws could not readily be applied, a more sober analysis of the history of the common law demonstrates that new business models and new technologies are regularly and inevitably incorporated into the law.

Section 230 was enacted in 1996 to address the narrow doctrinal problem created by Cubby v. CompuServe and Stratton Oakmont v. Prodigy. Over time, however, it hardened into a near-categorical immunity that bears little resemblance to the common-law duty-of-care framework it displaced. Online intermediaries today enjoy a liability shield that no comparable offline business has ever possessed. 

Section 230’s failure lies in ossifying into an immunity that excludes conduct it arguably should reach. The CFAA’s failure lies in stretching to reach conduct it was never designed to govern. But the underlying mechanism is the same: A statute drafted for a particular technological moment imposes categories tailored to that moment, while the common-law principles it displaced lose the opportunity to evolve alongside the technology.

The Sherman Act: Blessedly Vague 

Not every statute touching a fast-moving industry produces this kind of strain. The clearest counterexample is the Sherman Antitrust Act, which, as one of us has argued recently, “is best viewed as a modest statutory extension of the common law.” 

Read literally, the Sherman Act is almost comically terse. But its brevity proved to be a virtue. Its vague language and standards-based structure lent themselves to a common law-like process of judicial interpretation. Over time, courts developed the rule of reason, the per se / rule-of-reason distinction, and eventually the consumer-welfare standard. 

The result is a 136-year-old statute that has survived repeated technological revolutions—from railroads and broadcasting to software platforms and digital markets—because courts have continually adapted its principles to new facts. 

It did not have to turn out that way, and the contrast with more prescriptive competition statutes is instructive. The Robinson-Patman Act in the United States and the European Union’s Digital Markets Act (DMA) both reflect a more top-down approach. The DMA, for example, “applies per se rules to broad swathes of conduct in so-called digital markets,” with “no scope for effects analysis or procompetitive justifications.” 

That is the opposite of the common-law method. It assumes legislators can identify in advance, across industries and technological contexts, which practices will prove anticompetitive. The resulting error costs are predictable and often substantial. 

And the political-economy costs may be worse. Antitrust, like any body of law governing significant commercial interests, attracts rent-seeking by firms hoping to hobble competitors through regulation rather than competition. 

The CFAA, even after Van Buren, sits closer to the DMA end of the spectrum than to the Sherman Act end. Its framework relies on categorical concepts rather than standards-based balancing. If agentic AI ultimately requires a statute at all, the better model would look less like the CFAA and more like the Sherman Act: principles-based, flexible, and designed to leave courts room to adapt the law incrementally. 

The political-economy point matters here as well. When Congress enacted the Sherman Act, organized interest groups had only limited ability to shape its implementation. The same cannot be said of agentic AI. The interested constituencies are already organized, already lobbying, and already shaping the proposals now circulating in Washington. 

The impulse to “just write a new statute” risks producing something closer to the DMA’s rigid framework than the Sherman Act’s relative durability. 

The Digital Millennium Copyright Act: A Useful Warning

The Digital Millennium Copyright Act (DMCA) is perhaps the most instructive example because it occupies the middle ground. 

Congress enacted the DMCA in 1998 to solve a problem that neither the common law nor private contracting could easily solve on their own: governing access to copyrighted works at internet scale. 

The transaction-cost problem is straightforward. Copyright is a property right ordinarily enforced through contracts. Rightsholders license works, users pay for access, and contracts define the terms. But where transactions become highly dispersed, low-value, and extraordinarily numerous, bilateral contracting becomes prohibitively expensive. 

Historically, markets responded by creating collective institutions. Performance-rights organizations such as ASCAP and BMI emerged to facilitate music licensing where individual negotiations would have been impractical. The DMCA represented Congress’s attempt to create a comparable framework by statute before market institutions had developed at sufficient scale. 

Viewed charitably, that logic maps closely onto the agentic-AI access problem. Amazon and Perplexity can, in theory, contract with one another. But bilateral negotiations between every platform and every developer of every agentic system are unlikely to scale. A statutory framework featuring identified agents, safe harbors, opt-outs, and knowledge-conditioned liability could, in principle, reduce those transaction costs. 

The problem is that the DMCA also illustrates how quickly statutory compromises can drift away from their original design. 

The provisions intended to protect rightsholders have gradually been narrowed through judicial interpretation (so judicial interpretation is not a panacea, either). Section 512’s notice-and-takedown framework was intended to provide a workable enforcement mechanism when contracting was impractical. Yet courts have often interpreted the statute to impose far less responsibility on intermediaries than Congress appears to have envisioned. 

Safe harbors intended as conditional protections for intermediaries that take meaningful steps against repeated infringement have, through cases such as Viacom v. YouTube and UMG v. Veoh, evolved into something approaching categorical immunity. Knowledge standards have narrowed. Repeat-infringer requirements have weakened. Decisions such as the 9th Circuit’s Lenz v. Universal have further limited enforcement by reading a fair-use precondition into the statute, despite the traditional rule that fair use operates as an affirmative defense. 

The result is a perpetual game of whack-a-mole. Rightsholders issue automated takedown notices against content that often reappears almost immediately. Meanwhile, the intermediaries whose conduct the statute was meant to regulate enjoy a liability regime increasingly disconnected from the common-law principles of intermediary responsibility that inspired it. 

The DMCA therefore serves both as a model and as a warning. Some aspects of its design closely track common-law principles. Notice-and-takedown resembles traditional abatement doctrines. The Section 512 safe harbors can be understood as a form of least-cost-avoider analysis. 

Its failures emerged when courts transformed those conditional protections into something resembling the broad immunity that Section 230 became. The underlying property interests became underprotected, and the duty-of-care framework gradually hollowed out. 

The implication for agentic AI is straightforward. Any statutory attempt to address the transaction-cost problem should preserve the conditional protections that make the framework work. Otherwise, the same interpretive forces that transformed Section 230 and weakened the DMCA may produce yet another technology statute whose practical operation bears little resemblance to its original design. 

The Market Is Already Solving the Problem

Meanwhile, a statute is necessary at all in this context only where private ordering is too costly. The most encouraging feature of the current moment, however, is that the market is not waiting for the litigation to end. 

Identified-agent regimes, negotiated-access arrangements, and commercial licensing agreements are emerging in real time. Amazon’s first-party agent, Buy for Me, identifies itself through a transparent user-agent token and honors opt-outs communicated through robots.txt and direct merchant requests. Reddit’s litigation against Perplexity has proceeded alongside Reddit’s own program of paid licensing agreements with major AI developers. Large publishers are negotiating direct deals with frontier-model companies. And robots.txt itself—the nearly 30-year-old industry standard through which web crawlers honor exclusion requests—demonstrates that workable private-ordering arrangements can emerge without statutory intervention. 

Whether the resulting equilibrium is optimal in every detail is a question better answered by the parties negotiating it than by legislators or courts. Disputes between platforms and developers of agentic systems involve exactly the kinds of interests that contract law routinely governs: server capacity, monetization, security, brand integrity, and obligations to merchants and advertisers. 

When private parties possess both the incentive and the technical capacity to negotiate access terms, the default presumption should be that they will do so. The legal system’s role is to enforce the resulting agreements, not to specify their contents in advance. 

None of this implies support for forced interoperability, mandatory disclosure, compulsory licensing, or treating platform “openness” as a regulatory objective. The absence of interoperability is rarely evidence of market failure. Forced interoperability often carries significant costs to security, product design, and firms’ ability to differentiate themselves—costs that its advocates routinely understate. 

Nor does the argument imply that Amazon must admit any agent that identifies itself. Quite the opposite. Treating this as a property-and-contract problem means preserving the platform’s right to say no. 

There is, however, a risk running in the opposite direction. The emerging private-ordering equilibrium is itself vulnerable to capture if Congress intervenes. Whatever statute Congress writes for agentic-AI access, incumbent AI developers and incumbent platforms will both seek to shape it to protect their existing positions.

The DMCA’s Section 512 notice-and-takedown regime, whatever its virtues, plainly reflects the interests and bargaining power of the incumbents of 1998. An AI-access statute enacted in 2026 would similarly reflect the interests and bargaining power of the incumbents of 2026, with predictably distortive consequences. 

The default position, unless and until Congress can devise a framework that supports rather than supplants common-law principles, should be to let parties negotiate, let markets adapt, and let courts resolve disputes using the property, contract, and tort doctrines we already possess. 

‘The Only Winning Move Is Not to Play’

For the 9th Circuit, the appropriate posture is restraint. If the panel concludes that Amazon has the stronger CFAA claim under existing doctrine—as it should—it ought to write narrowly. Any holding should turn on persistent circumvention of a deployed technical access control, not on a generalized theory of agentic access. Van Buren counsels exactly that approach. The court should resist the temptation to transform the CFAA into a general framework for governing AI agents’ interactions with commercial websites. 

The longer-run framework lies elsewhere: in property, contract, and trespass. 

For platforms and AI developers, the most productive path forward—private ordering—is, as noted, already emerging. Litigation will continue, and some of it will be necessary. But the equilibrium in this market will be shaped less by judicial opinions than by the agreements parties strike. The law’s modest contribution is to make property, contract, and trespass remedies predictable enough that those agreements become easier to negotiate. 

For Congress, the lesson is one of humility and timing. The CFAA and Section 230 are cautionary tales. The DMCA, the closest analog to a common-law-supportive statutory framework, has generated enough difficulties of its own to warrant skepticism. The Sherman Act is the rare success story, but its success depended in part on the political conditions of 1890—conditions that plainly do not exist in 2026. 

The default should be restraint. 

Amazon v. Perplexity is useful not because it presents a particularly difficult dispute, but because it exposes a recurring institutional temptation. Amazon’s interests are legitimate. Perplexity’s conduct, on the current record, is difficult to defend. The doctrinal stretch required to reach the right result under the CFAA is small enough that the court will likely make it. 

But that does not mean the statute is the right tool for the job. 

The broader lesson is that not every technological development requires a new legal framework, and not every legal problem requires a new statute. The common law has spent centuries adapting old principles to new facts. Agentic AI is new. The underlying questions of property, consent, agency, and trespass are not.

Let platforms exclude. Let agents identify themselves. Let parties negotiate. Let courts enforce the bargains they reach.

The law has seen this movie before.

The post WarGames, Shopping Bots, and the Statute Trap: The CFAA and Amazon v Perplexity appeared first on Truth on the Market.

That is the question Brussels would rather not answer after Apple announced that its new Siri AI features will not ship on iPhones and iPads in the European Union. The European Commission says Apple made a free choice. Apple’s actual choice was between opening the iPhone in ways that could break the privacy-and-security model its customers buy, shipping a product that would invite enforcement, or not shipping at all. 

Call it innovation by ultimatum. 

At Tuesday’s midday press briefing, European Commission spokesperson Thomas Regnier offered Brussels’ official explanation: 

The decision not to allow Siri AI in the EU is Apple’s and Apple’s only… Absolutely nothing in the DMA prohibits Apple from introducing new products in the EU… What Apple is, however, not allowed to do is to close the market. It’s not for them to choose which AI tools our EU citizens get to use or not… EU law is non-negotiable.

Regnier capped the intervention with an analogy: the Commission grants no exemptions, just as a police officer does not exempt a driver from the speed limit. 

The problem is that nearly every clause of his statement gets the economics backward. 

The Commission’s framing is that Apple faced a free choice and chose spite. The reality is closer to a catch-22. Under the Commission’s interpretation of the Digital Markets Act (DMA), Apple could open its operating system in ways that compromise the security architecture its customers pay for—and that European Union privacy law itself demands—or it could preserve that architecture and ship nothing. 

Apple chose the only option that was both lawful and commercially rational: it withheld the product. When a regulatory regime is structured so that the only safe harbor is nonparticipation, the claim that “nothing prohibits you from introducing new products” is technically true but substantively empty. 

To see why, it helps to recall what actually happened. It also requires asking a question the Commission has studiously avoided: If the open, deeply interoperable, agent-accessible operating system Brussels demands is both commercially viable and consistent with the rest of EU law, why has no one—on any platform, anywhere in the world—ever built one? 

The Architecture of a Catch-22

Apple unveiled Siri AI on June 8 at the Worldwide Developers Conference (WWDC): a rebuilt assistant, powered by Apple Intelligence, capable of carrying out multi-step conversations and acting autonomously across a user’s device. It can search messages and email, complete bookings, make payments, and edit files. In the same announcement, Apple said the feature would not ship on iOS 27 or iPadOS 27 in the European Union, and offered no timeline for its arrival. 

The same assistant will reach European users on Mac and Apple Vision—platforms the DMA does not designate as gatekeepers. That detail alone answers one question: the binding constraint is the regulation, not the technology. 

The sticking point is the DMA’s interoperability obligations. Under what Apple calls an “extreme interpretation” of those rules—but one the Commission appears to have embraced—shipping Siri AI in the EU would require giving rival AI agents equivalent access to the device. Those agents would be able to read messages, control apps, make purchases, and edit files autonomously, without Apple’s intermediating safeguards. 

Apple says it spent 18 months proposing alternatives. Those included a “Trusted System Agent” architecture that would have extended similar capabilities to third-party assistants through a mediated layer, as well as a phased rollout. The Commission rejected those proposals, insisting on openness without the accompanying safeguards. 

Now consider how Thomas Regnier described that history. Apple, he said, was “simply unable to develop interoperability solutions that meet essential EU privacy and security standards” and therefore requested an exemption instead. 

That sentence is worth reading twice, because it may be the most revealing thing any European Union official has said about this dispute. 

The Commission’s own spokesperson is effectively telling us that opening the iPhone to third-party AI agents in the manner the DMA requires could not be done—at least not by the company with arguably the greatest engineering resources and the strongest commercial incentive on Earth to do so—without violating the EU’s own privacy and security standards. 

Regnier presents that conclusion as an indictment of Apple’s diligence. It reads instead as a confession about the legal architecture. 

How to Ban a Product Without Banning It

The decision not to allow Siri AI in the EU is Apple’s and Apple’s only… Absolutely nothing in the DMA prohibits Apple from introducing new products in the EU…

Both halves of that statement are technically true. Together, they describe a menu of options that effectively left Apple with no meaningful choice and no viable path to introducing the product. 

Option 1: Violate the GDPR

Hand autonomous, system-level agents from any third party the keys to users’ messages, payments, photos, and files. 

Quite apart from the commercial damage—Apple’s entire competitive strategy rests on being the privacy-and-security platform, a distinction that even its courtroom adversaries have acknowledged as genuine—this option collides head-on with the European Union’s other flagship digital statute.