The researcher conducted a large-scale experiment involving over 500 professionals across consulting, data analysis, and management. Participants completed authentic workplace tasks—from revising financial reports to conducting statistical analyses—using one of 13 different large language models (LLMs) spanning several generations of AI capability. The results establish what the paper calls “scaling laws for economic impacts”: quantifiable relationships between a model’s computational power and a worker’s productivity.

The headline figures are pretty striking. Each year of AI progress reduces task completion time by roughly 8%. Access to any AI model increased workers’ earnings per minute by 81% and raised the quality of their output by a third of a standard deviation. When accounting for performance bonuses, total earnings jumped 146%. If we extrapolate these trends forward, the researcher estimates that the continued scaling of AI could boost American productivity by approximately 20% over the next decade.

The great divide

Yet these aggregate numbers mask critical heterogeneity. The productivity gains were dramatically larger for what the researcher calls “non-agentic” tasks—analytical work like hypothesis testing or report writing that requires thought but little interaction with external tools. For these activities, AI assistance increased total earnings by $1.58 per minute.

But for “agentic” tasks requiring multi-step planning and tool use—creating spreadsheets, extracting data from multiple documents, running statistical tests—the gain fell to just $0.34 per minute, a difference that proved statistically significant.

This divergence suggests that whilst current AI systems are rapidly commoditising analytical cognition, they remain considerably less effective at procedural work requiring careful orchestration of multiple steps. The finding carries important implications for labour markets: jobs centred on pure analysis may face greater displacement than those requiring complex project management or hands-on execution.

The human ceiling

Perhaps the study’s most striking discovery concerns output quality. When the researcher tested the AI models in isolation—having them complete tasks without human involvement—they found robust scaling laws. As training compute increased tenfold, the quality of autonomous AI output improved by half a point on a seven-point scale. The most capable models achieved grades exceeding six out of seven, well above the human-only average of 3.5.

But when humans worked with these models, something peculiar happened: the scaling law disappeared entirely. Whether participants used weak early models or powerful frontier systems, the final output quality remained stubbornly fixed at roughly 4.3 out of seven.

For weaker models, human intervention improved results—workers successfully refined imperfect drafts. For intermediate models, humans added nothing. For the strongest models, which autonomously produced work scoring six or above, human involvement actually degraded quality, pulling it back down to the 4.3 average.

This represents what might be called the “satisficing ceiling”: humans appear to have an internal quality threshold they’re satisfied with, rather than maximising the tool’s potential. They edit AI-generated excellence into mediocrity. The phenomenon recalls earlier research on calculator use—people checking obviously correct answers because they don’t trust the machine, or because “showing their work” feels necessary.

Compute versus cleverness

The research also sheds light on the sources of AI progress. By comparing models released at different times but trained with similar amounts of computing power, the researcher decomposed productivity gains into two components: pure scaling of computational resources, and algorithmic improvements like better training methods or data quality.

The results suggest that roughly 56% of recent productivity improvements stem from increased compute, whilst 44% come from algorithmic innovation. This has practical implications. If progress depended entirely on ever-larger training runs, AI advancement would be constrained by the massive costs and energy requirements of frontier models. But the substantial role of algorithmic progress offers a more sustainable path forward.

Caveats and concerns

The study has limitations, which the author acknowledges. Participants used standard chatbot interfaces with restricted tool access, potentially underestimating what today’s more sophisticated AI assistants can achieve on agentic tasks. The tasks, whilst realistic, were relatively short-horizon activities taking under an hour—quite different from the multi-day projects common in professional settings. And the experiment captures individual productivity, not team dynamics or broader organisational effects.

There are also questions about how applicable it is to the various settings usually found in a typical workplace. For instance, the participants, while professionally experienced (79% had three or more years in their field), were recruited through online platforms rather than sampled from traditional firms.

The author also acknowledges that the experience of AI across the volunteers may also be different to that of the wider workforce. What’s more, the aggregate productivity projections assume that inference costs remain as low in the future as they are today, which may not be the case as LLM operators aim to recoup their huge costs.

The productivity paradox ahead

Still, the research represents a methodological advance. Where previous studies compared “AI versus no AI,” this experiment traced how productivity responds to continuous improvements in model capability. It offers economics what physics has long possessed: empirical scaling laws that allow forecasting based on measurable inputs.

The apparent human ceiling on quality remains puzzling, however. If employees remain reluctant to exploit the full capabilities of AI, then the huge investments the tech companies are making in new capabilities might yield diminishing returns in terms of productivity. The bottleneck is less the technology and what it’s capable of, as it is humans and what we’re willing to do with the technology.

The author suggests that two possible futures may result if this continues. One would see improved interfaces and better training helping employees better capitalize on what AI can do, therefore boosting productivity. In the other, we continue getting by with good enough, and productivity plateaus even as AI becomes more and more capable.

The study also shows where AI can produce the best value today, with analytical tasks, such as synthesizing information and reading documents, particularly good tasks for AI to help us with. When tasks are more complex and multi-layered, however, the gains are far less substantial. The dream of AI agents performing complex and intricate workflows remains a way off.

The research reminds us that while the tech industry invests billions (even trillions) in better AI, the path between the capabilities of the technology and better productivity has to pass through a human bottleneck that requires human solutions to address. Unlocking AI’s full potential may require not just better algorithms, but a better understanding of how humans and machines can genuinely complement each other’s strengths.

The researchers are careful to distinguish proximity from participation. Simply glimpsing greenery from a window, or residing in a neighbourhood with a respectable tree canopy, is not enough. The real benefits accrue only when people venture outside and actively spend time in such spaces—ideally every day, or at least several times a week.

Access to nature

Drawing on a survey of 1,361 Floridians aged 45 and over, the study examines how stress correlates with demographics, health, and access to urban forests. Respondents cited familiar anxieties, notably financial insecurity and poor health. Against this backdrop, the researchers tested a popular heuristic known as the “3-30-300 rule”: at least three trees visible from one’s home, a neighbourhood canopy cover of 30%, and access to a recreational green space within 300 metres.

Nearby trees helped, but they were no substitute for immersion. The lowest stress levels were reported by those who spent time daily on tree-lined streets or in parks. Conversely, limited access to green spaces—and infrequent visits—was associated with significantly higher stress.

The findings strengthen the case for planting more trees in cities. Their value is often framed in terms of shade or visual appeal, the researchers note. Less appreciated are the mental-health dividends. As urban density increases and development squeezes out green space, those dividends may become harder to ignore—and more costly to forgo.

The study focuses on how founders respond to visible participants in online discussions. Those replies, it argues, perform a dual function. They address the immediate interlocutor, but they also act as a public signal to silent observers who are forming judgements about the idea’s credibility and relevance. The researcher terms this strategy “discourse broadening”: replies that move beyond a narrow back-and-forth and touch on a wider set of issues, perspectives, or stakeholder concerns.

Increasing support

Using data from 9,746 product launches by 8,489 entrepreneurial teams on Product Hunt, the analysis shows that such broadening can significantly increase support from the silent majority. This matters because silent users—who rarely comment but do decide what to upvote, share, or ignore—ultimately determine which ideas gain visibility and legitimacy.

The mechanism is indirect. Entrepreneurs do not persuade silent audiences through targeted argument, but by modelling inclusiveness in public exchanges. As the author notes, founders often believe their task is to convince those they engage with most directly. In fact, the harder task is mobilising those who remain quiet. How one handles a few voices in public can shape how everyone else interprets the idea.

There are limits, however. The relationship is not linear. Some broadening helps; too much obscures the core message and breeds uncertainty. The optimal degree depends on how representative the vocal commenters are of the wider audience. Where they already mirror the crowd, extensive broadening adds little. Where they do not, it can be decisive.

Practical lessons

For founders and venture communicators, the lesson is practical. Effective online engagement is not merely reactive. It involves acknowledging a commenter’s point while briefly extending the discussion to adjacent concerns—such as usability, outcomes, or wider social relevance—so that different observers can see why the idea might matter to them.

In diverse audiences, this calibrated approach works best. It creates multiple entry points without diluting the narrative. More broadly, the findings highlight how influence in digital markets flows through visible conversations, and why managing them carefully has become a strategic task for entrepreneurs seeking to be noticed by those who never speak.

Drawing on data from nearly 100,000 respondents across 66 countries, the researchers find that a sense of autonomy—defined as the freedom of choice and control over how one’s life turns out—is associated with higher well-being everywhere. But the strength of that association varies markedly. It is strongest in wealthy, individualistic countries, such as the Nordic states, and weaker in poorer ones, where more basic concerns tend to crowd out the pursuit of self-direction.

Solving the dispute

This resolves a long-running academic dispute. One camp has argued that autonomy is a fundamental psychological need and should therefore matter equally across cultures. Another has maintained that it is a value prized mainly in affluent, individualist societies. The new evidence suggests that both views capture part of the truth.

To reach this conclusion, the team combined measures of happiness and life satisfaction from the World Values Survey with national income data, using per-capita GDP adjusted for purchasing power. They then examined how national wealth shapes the autonomy–well-being relationship. Measuring culture proved trickier. Rather than relying on blunt or inconsistent proxies for individualism, the researchers used a newer tool, the Global Collectivism Index, which focuses on observable behaviors such as shared households and transportation use.

The policy implications are pragmatic rather than ideological. In very poor countries, raising incomes—ideally in a broadly shared way—remains the most effective route to improving well-being. As nations grow richer, however, the marginal gains from wealth diminish, and autonomy begins to matter more. Governments, employers, and institutions would do well to take note.

Autonomy, after all, is not confined to constitutional rights or the absence of repression. It is also shaped by everyday experiences, particularly at work. If it is a basic human need, as the evidence suggests, then supporting it requires attention not just to laws and markets, but to the design of the societies in which people live.

Changing preferences

Negative screening has long dominated, largely because it is simple and legally uncontroversial. But retail investors, the study finds, now show a growing preference for positive screening, which constructs portfolios around firms with high ESG scores. Institutional investors, typically more sophisticated, have been ahead of the curve. Though positive screening is more complex, the researchers argue that it offers superior protection against downside risks.

Surveyed investment managers concede that while negative screening remains common practice, it may come at a financial cost. By contrast, positive screening broadens the investable universe, improving diversification and, with it, risk-adjusted returns. Firms that score well on ESG metrics tend, the study notes, to be less exposed to shocks such as extreme weather or regulatory backlash, making them steadier performers during downturns.

The data bear this out. ESG-focused portfolios appear less volatile and more resilient in periods of market stress. Both retail and institutional investors are already voting with their wallets, allocating substantial sums to funds that employ positive ESG screening.

For fund managers, the implications are clear. Negative screening may minimise legal headaches, but it no longer reflects investor preferences—or their appetite for risk management. As uncertainty becomes a permanent feature of markets, selecting the best corporate citizens may prove not just virtuous, but prudent.

The pessimistic view of AI rests on a simple intuition: if machines can do what workers do, workers become redundant. Yet this misses a crucial dimension of how technology affects labour markets. The authors identify three distinct channels. Augmentation makes workers more productive at tasks they already perform. Automation allows machines to complete tasks autonomously. But the third channel, which the authors call simplification, has received less attention. It captures how AI can reduce the skill level needed to perform a task—think of how spell-checkers made good writing accessible to those with poor spelling, or how GPS navigation reduced the expertise required to drive a taxi.

How we choose

To quantify these effects, the researchers built a model that tracks how workers with different skill combinations choose occupations, perform tasks, and accumulate expertise over their careers. They estimated the model using decades of data on American workers’ wages, job choices, and multidimensional skills. The framework allows them to predict how any technology that changes task requirements will ripple through the economy as workers retrain, switch jobs, and wages adjust.

Measuring AI’s capabilities required creativity. The authors used large language models to assess how AI affects nearly 20,000 tasks across the economy. They estimate that generative AI can fully automate 22% of tasks, save workers an average of 20% of their time on others, and reduce average skill requirements by 18%. These estimates align reasonably well with experimental studies of AI in workplaces, from management consulting to customer service.

The model’s central finding is that AI’s impact on inequality is “fully driven by simplification.” Without this channel, AI would raise average wages but slightly increase inequality, as in standard models of automation. Simplification, by contrast, enables lower-skilled workers to compete for jobs previously reserved for their better-educated peers. A radiologist’s job becomes easier; so does a management consultant’s. This compression of skill requirements translates directly into compressed wages.

A double-edged sword

Yet simplification is a double-edged sword for individual occupations. Consider management analysts, who rank in the 97th percentile for how much AI simplifies their work. The model predicts their employment will fall by 4% even as wages stagnate—well below the 21% average gain. Why? As tasks become easier, more workers can perform them competently. Increased competition suppresses wages even as productivity rises. For some occupations, the jobs that experience the largest employment losses also see the highest wage increases, a “negative relationship between wage and employment effects” that arises precisely because simplification enlarges the pool of qualified workers.

The distributional effects are striking. Workers in the bottom quarter of the wage distribution see gains of up to 40%, while those at the very top see almost no benefit. The returns to different skills shift dramatically. Verbal skills, once highly valued, become less so; mathematical ability holds its ground. Ex-ante welfare gains—measured as the permanent wage increase that would make workers equally well off—range from 26% to 34% for most workers, with the largest gains accruing to the least skilled.

Early evidence suggests the model’s predictions may already be materializing. Using an event-study design around ChatGPT’s November 2022 release, the authors find that occupations predicted to benefit from AI have indeed begun to see relative employment gains, with roughly 8% of the predicted long-run effects visible by mid-2025. University students, too, appear to be responding: enrollment in majors the model predicts will benefit from AI—such as atmospheric sciences and engineering—has increased relative to those predicted to struggle, like theology and French.

Not all occupations fare equally. Radiologists, despite dire predictions that AI would render them obsolete, are forecast to see their wage bill rise 42%—driven by strong employment growth that more than offsets below-average wage gains. The reason: most of their tasks cannot be fully automated, and the automatable ones tend to be the least skill-intensive. Office and administrative support workers, by contrast, face grim prospects, with wage bills predicted to fall in absolute terms.

Crossing over

The analysis assumes AI remains limited to processing digital information. But the authors also explore a scenario in which AI gains physical capabilities through robots and autonomous vehicles. Average wages would rise by 39% rather than 21%. Yet occupations requiring manual skills—food preparation, farming, production work—would flip from winners to losers. The returns to mathematical skills would rise further while other dimensions decline more sharply.

The paper’s methodology represents an advance for economists trying to peer into the future. Previous models of technological change have struggled to predict how workers will adapt because they treat skills as fixed and ignore how comparative advantage shapes who does which jobs. By integrating multidimensional skills, dynamic learning, and task-based production into a single framework, the researchers offer a toolkit for evaluating any prospective technology’s labour-market effects.

Of course, all models simplify reality. This one assumes workers care only about wages and amenities, ignoring non-pecuniary concerns like job security or meaning. It treats technology as exogenous rather than directed by profit-seeking firms. And it focuses solely on labour income, overlooking how AI might concentrate capital income or reshape entrepreneurship.

Still, the core insight is robust: when technology makes tasks easier rather than merely faster or automated, it can reduce inequality by democratizing access to high-skill work. If the authors are right, the AI revolution may prove less dystopian than many fear. The future of work looks not like mass unemployment, but like a world where more people can do more kinds of jobs—even if some prestigious occupations find their expertise less scarce than before.

The study finds a striking overlap between places where environmental burdens cluster and where vulnerable populations live. Census tracts facing multiple hazards—ranging from toxic air pollution to wastewater discharge—are disproportionately home to communities of colour and to people with limited English proficiency. Indeed, the presence of such burdens is a stronger predictor of those communities than of any other socio-demographic group.

Deliberate acts

That pattern is no accident. The correlations are far stronger for race and ethnicity than for income or other measures of disadvantage, reflecting the legacy of decades of public policy that segregated neighbourhoods and concentrated risk—redlining among them.

Most research has treated environmental harms in isolation, largely because the data allowed little else. People, however, do not experience pollution one variable at a time. They live amid overlapping hazards and social stressors that compound one another. The Michigan researchers address this gap by examining up to 13 environmental burdens across more than 84,000 census tracts in all 50 states and Puerto Rico, while also accounting for their severity.

Such work has practical implications. New Jersey recently became the first state to require regulators to deny permits in areas already overburdened by cumulative impacts, using its own environmental-justice mapping tool. Studies like this one could help other states follow suit, making assessment tools both more rigorous and more widely available.

Data is key

The data revolution that made the study possible has been gathering pace for a decade. Federal and state agencies have gradually expanded access to environmental and demographic information, culminating in EJScreen, a national assessment tool launched roughly ten years ago. By combining EJScreen with census data, the researchers could map where hazards accumulate and who lives among them—revealing the geography of inequality with unusual clarity.

Even so, gaps remain. The health and economic consequences of every environmental burden are not yet fully documented. But, as the authors note, it is a reasonable assumption that as the number and intensity of such burdens rise, so too does the harm.

Federal support for this line of inquiry is now retreating. The current administration has effectively shuttered EJScreen. Yet the work continues at the state and local levels. As long as environmental inequities persist, researchers, policymakers, and communities are unlikely to look away. The attention will fade only when the disparities do.

The study looks beyond pay packets to attitudes. What matters, the authors argue, is not the absolute size of an income but what money represents within a partnership: a marker of achievement, a symbol of status, or merely a means to an end. Those beliefs, it turns out, do not stay at home when one partner goes to work.

Winning the bread

Men’s sense of fulfilment in their jobs is closely bound up with their partners’ views on money. That is because earnings are not generated in a vacuum but within households. Evaluating the role of income, the researchers contend, requires looking at the couple as a unit rather than the worker in isolation.

When partners share similar financial values, husbands tend to feel more satisfied with their work. When they diverge, that contentment ebbs. Curiously, the lowest levels of fulfilment appear not when couples hold starkly opposing views, but when their positions are only partially misaligned. This “messy middle”, as the researchers describe it, breeds uncertainty, making it harder for men to take confidence from their professional lives.

The pattern is far weaker for women. Their sense of workplace fulfilment appears largely unaffected by their partners’ attitudes to money. The authors attribute this asymmetry to the lingering power of cultural expectations. Even as dual-income households have become the norm, ideas about the male breadwinner and the female carer remain deeply embedded—particularly in men’s own self-conceptions.

Identity still matters

For many men, the pressure to live up to a breadwinner identity persists in a labour market that no longer neatly supports it. That identity, the study suggests, is remarkably sticky. As a result, alignment—or misalignment—with a partner’s financial values can materially shape how men feel about their work.

The practical implication is blunt. Career choices matter, but so do domestic ones. In an era when most workers share both bills and beliefs with a partner, harmony over money may prove almost as important to job satisfaction as the job itself.

It nicely captures the sense that things like art and writing are inherently human activities that should be free from AI encroachment. Of course, that hasn’t been the case, and AI slop was selected as the word of the year by Merriam-Webster as a reflection of the huge quantity of AI-generated content hitting the web.

To write or not to write

For many of those generating such content, the decision is less about artistic, creative, or intellectual merit than it is about the content being good enough for their commercial goals. Such people will no doubt be buoyed by a new study that suggests that artificial intelligence can now replicate an author’s style so convincingly that even trained literary experts struggle to tell the difference. The implications for creativity, commerce, and copyright law are profound.

The study was conducted by researchers from the University of Michigan, Stony Brook University, and Columbia University, who created a series of experiments that should trouble anyone who earns their living by the pen. When AI systems were given simple instructions to imitate an author’s voice, which is a technique called “in-context prompting,” experts trained in Master of Fine Arts programs could usually spot the forgery. Ordinary readers, however, often preferred the machine’s output to that written by humans.

Fine-tuned prose

The real shock came when the researchers moved to “fine-tuning”, a method that trains the AI model on an author’s complete works. At that point, even the experts’ defences crumbled. They frequently rated the AI’s writing as superior both in stylistic fidelity and overall quality.

The economics are stark. Training an AI to mimic an author’s style costs roughly $81. Hiring a professional writer to produce the same quantity of prose costs about $25,000. That represents a 99.7% reduction in cost—the sort of efficiency gain that tends to reshape industries overnight.

The study reveals a telling divergence between expert and lay preferences. MFA-trained readers recoiled at AI prose generated through simple prompting, detecting what the researchers refer to as  “performatively literary” writing, which usually involves over-explained emotions, stock metaphors, and an unearned smoothness.

Ordinary readers, by contrast, valued clarity and flow—qualities at which AI excels. Fine-tuning closed this gap by teaching the model to absorb an author’s recurring rhythms and lexical habits rather than relying on generic “literary” patterns.

Hard to mimic

Not all voices proved equally susceptible to imitation. Tony Tulathimutte, an author known for transgressive humour and sharp tonal shifts, remained elusive even after fine-tuning. The model could produce fluent prose that laypeople rated highly, but it struggled to capture his distinctive voice. This suggests that whilst AI can master many stylistic elements, truly idiosyncratic voices retain some measure of protection, at least for now.

The legal ramifications are already playing out in American courts. Lawsuits, including Kadrey v Meta and Bartz v Anthropic are testing how copyright law applies to AI training and output. The study strengthens the plaintiffs’ case on a crucial point: the fourth fair-use factor, which considers the effect on the market for copyrighted works. If lay readers already prefer AI-generated prose, and fine-tuning can convince even experts, then the threat of market substitution becomes concrete rather than hypothetical. Author-specific fine-tuning designed to produce style-emulating outputs looks harder to defend as fair use, particularly when marketed as a replacement for human writers.

The ethical questions are perhaps more vexing than the legal ones. Authors rarely consent to having their works used to train systems that can imitate them. Compensation mechanisms barely exist. And whilst general principles about disclosure and provenance sound reasonable, enforceable mechanisms remain elusive. What obligations do AI companies have when they make style imitation both cheap and effective?

What now?

For writers and writing programmes, denial is not an option. Several MFA-trained experts interviewed after the study expressed unease. “I thought I could tell what AI was, and now I’m not so sure,” one said. Creative-writing programmes must now teach students not just how to write well, but how to develop voices distinctive enough to resist mechanical replication. They should also address practical questions: when does AI assistance aid exploration, and when does it flatten voice into something generic?

The ghost in the machine, it turns out, can learn to write like almost anyone. It’s almost certain that AI slop isn’t being rejected because it’s terribly written, however, and therefore improving the quality of the prose might have a limited impact on its acceptance by readers. The issue is more that we simply don’t want to read what AI produces, as the act of reading is, as the original meme suggests, an inherently human process.

Drawing on Finland’s unusually rich population registers, the researchers find that senior corporate executives are far more likely to commit financial crimes if their parents had previously been convicted of similar offences. The association is strongest when parental crimes were serious enough to result in prison sentences. The imprint of early life, it seems, does not fade easily—even in the upper reaches of corporate leadership.

Behind the crime

The study examines financial crime convictions of chief executives and board members of Finnish limited-liability companies while they were in office. Using judicial records from Statistics Finland, it tracks nearly 76,000 executives across more than 64,000 firms between 1995 and 2019. The offences range from accounting fraud and tax crimes to other forms of economic misconduct, and include only cases that resulted in formal charges and convictions.

Family influence extends beyond parents. Executives convicted of financial crimes are disproportionately likely to have spouses with similar convictions, and the probability of misconduct rises with the length of the relationship. Geography, too, matters. Those who grew up in areas with above-average levels of financial crime are more prone to offend later in life.

These patterns are not easily dismissed as the by-product of shared family businesses or low socioeconomic status. The researchers explicitly control for such factors and find that they do not explain the results. Nor do the findings imply a simple chain of cause and effect. Rather, they point to the power of social learning: norms absorbed in childhood and reinforced in close relationships appear to shape behaviour decades later.

Sobering implications

The implications are sobering. Financial crimes committed by top managers rarely harm only their perpetrators. They destroy jobs, erode tax revenues, and wipe out invested capital, often affecting thousands of people. That such damage may be linked to deeply ingrained behavioural patterns complicates the task of prevention.

The authors argue that effective efforts to curb financial crime must therefore go beyond tighter regulation and sharper incentives. They must also address the need to “unlearn” norms acquired early in life. White-collar crime has long been under-researched, and corporate responsibility has only recently become a staple of business-school curricula. By tracing intergenerational links in executive misconduct for the first time, this study offers a new lens on why some leaders break the rules—and why culture, not just controls, still matters at the top.