Using comprehensive Swedish data, the study examines what happens when firms expand far faster than their historical growth trajectory would predict. Such bursts of hiring, it finds, compress decision-making and encourage reliance on cognitive shortcuts. Under pressure, founders are more likely to fall back on familiar stereotypes, shaping both who is hired and who is promoted.

Frozen out

The analysis covers more than 31,000 ventures founded between 2004 and 2018. In male-led startups, scaling reduces the likelihood of hiring a woman by roughly 18%, and cuts the odds of appointing a woman to a managerial role by 22%. That these effects appear in Sweden—a country with strong norms and institutions around gender equality—makes them harder to dismiss as cultural artefacts.

There is, however, an important mitigating factor. Founders with formal training in human resources practices behave differently. In startups led by such founders, scaling increases the odds of hiring women by over 30%, and raises the probability of appointing women to management by around 14%. Familiarity with structured hiring processes appears to anchor decisions in process rather than instinct.

Experience in firms with established HR systems also helps, though more modestly. It improves women’s chances of being hired as companies scale, but does little to close the gap in managerial appointments.

A consistent pattern

The biases identified are not confined to male founders. Solo female-led ventures show similar, if weaker, patterns under rapid growth. Nor are female-dominated industries immune: scaling boosts women’s hiring into non-managerial roles, yet still diminishes their chances of reaching the top.

The lesson is not that founders are malicious, but that time pressure is corrosive. When growth accelerates, cognitive bias asserts itself across the board. The study’s implication is prosaic but powerful: getting basic HR disciplines in place early may be one of the most effective—and overlooked—ways to preserve fairness when success comes quickly.

“Responsible AI” currently operates as an aspiration without accountability. Companies claim compliance while practicing widely varying behaviors. There are no enforcement mechanisms and no consequences for failure.

A chaotic regulatory environment is particularly harmful to women and marginalized groups. It emboldens companies to pursue profit, safe in the knowledge that they can shop around for the most favorable jurisdictions and avoid the various activities, such as bias audits, that might make AI more equitable.

Public infrastructure

As AI’s importance has grown, conversations around digital sovereignty have focused on the apparent reliance on foreign private systems.

Various open-source models, such as Llama and Mistral, show that foundation models don’t need to be built from scratch. Indeed, China’s DeepSeek showed that competitive models can be built quickly and cost-effectively.

As a result, it’s clear that governments can develop models that are ethical without spending the vast sums the tech giants are spending. It’s not about matching all of the features of GPT-5. It’s about ensuring that public-interest AI is available for all for the most critical applications.

Fighting back

Open source has long been the David to Big Tech’s Goliath. From Linux to Kubernetes, collaborative development has powered the digital revolution, offering a counterweight to corporate control. Now, artificial intelligence is testing whether this model can survive contact with reality. A new paper by Joshua Tan, Nicholas Vincent, and colleagues argues that it cannot—at least, not without help.

The authors make an uncomfortable claim: open-source AI, as currently practised, will not democratise access to intelligent machines. Unlike software libraries that can run on ordinary computers, frontier AI models demand extraordinary resources. Training them requires thousands of graphics processing units running for months. Operating them at scale needs constant access to expensive infrastructure. Fine-tuning them for practical use involves proprietary datasets and closed techniques. The weights may be open, but the systems remain stubbornly closed.

This matters because the gap between availability and usability is widening. Meta’s Llama models, though freely downloadable, see just 6m monthly downloads on Hugging Face—a pittance compared with OpenAI’s consumer adoption. Community efforts like EleutherAI’s Pythia manage fewer than 1m. The strongest “open” models today all come from well-capitalised firms. Even when activists succeed in training models, they often lack the infrastructure to make them useful. Releasing model weights without inference, tooling, and interfaces is rather like publishing sheet music without orchestras to perform it.

Public AI

The paper proposes “public AI” as a remedy: government-funded infrastructure for deploying, maintaining, and governing AI systems. This is not mere theory. Sweden has built GPT-SW3 to serve Nordic languages poorly handled by ChatGPT. America’s National Deep Inference Fabric aims to provide democratic access to open models. India, Canada, China, and the United Arab Emirates are all experimenting with state involvement in AI development.

Critics will bristle. Markets, they argue, are working splendidly. OpenAI and Meta offer advanced models cheaply or for free. Private labs iterate faster than any bureaucracy could. Why burden innovation with public spending and governance?

The authors’ response is that access is not governance. Commercial generosity can be withdrawn at any time. Meta’s Llama licence, for instance, contains restrictive terms and can be revoked unilaterally. More fundamentally, open source has always relied on hidden subsidies—employment by tech firms, cloud credits, and academic positions. For AI, these subsidies are breaking down. Training costs are soaring. The community contributions that once built shared infrastructure now risk fuelling proprietary systems that appropriate the value.

A variety of approaches

Public AI need not mean nationalisation. The paper envisions various models: governments contracting private labs, academic networks collaborating on development, or direct public provision of AI services. The goal is to ensure that open models remain accessible, trustworthy, and competitive—not to replace private innovation but to complement it.

Sceptics worry about inefficiency and capture. Public-sector technology projects have a patchy record. Yet well-governed public institutions have delivered GPS, the Internet, and the Hubble Space Telescope. The European Organisation for Nuclear Research (CERN) shows that public science can resist capture and reward quality. The real risk may lie in failing to provide an alternative to concentrated private power.

For researchers, public AI could democratize frontier experimentation. Shared infrastructure would let scientists access model internals and query data without hosting their own hardware. For domains like healthcare and law, it would enable experts to adapt systems without relying on corporate APIs. For governments, it offers digital sovereignty and a stake in how powerful systems develop.

Open source must evolve

The machine-learning community, the authors argue, should not conflate open source with public good. Models are not like traditional code. They are “impure public goods”—non-rival in theory but requiring private complements like compute and energy to be useful. As AI systems grow more complex, the gap between downloadable weights and usable tools will only widen.

If open source is to fulfil its democratic promise in the age of AI, it must evolve. That means recognising its material constraints and building public infrastructure to overcome them. The romantic view of open collaboration must give way to a pragmatic one: AI as public infrastructure, sustained by institutions accountable to more than shareholders. Whether governments have the will to build such systems remains uncertain. What seems clear is that leaving AI entirely to the market will not suffice.

Research from the University of New South Wales shows that both partners working from home can be a frustrating experience for both parties, although, as is often the case, a heavier toll is placed on the woman.

Always on

A lot of the tension arises from the intrusion of work into family time as a result of our digital devices, causing constant interruptions. It’s a concept the researchers refer to as “ICT permeability,” which is when our digital devices erode the line between work and home life.

This is a well-documented issue when we work from home, but the researchers show that it’s even more problematic when both partners work from home.

Their research, which tracked 117 dual-earner couples over ten working days, reveals a paradox at the heart of remote work: the same technology that strains family relationships also boosts job performance.

Rather than relying on one-off surveys, the researchers used “experience sampling,” asking participants to respond to questionnaires three times daily for a fortnight. This approach captures the granular reality of work-from-home life—the Zoom call that runs into dinner time, the email answered whilst helping children with homework, the late-night Slack message that cannot wait until morning.

Not universally harmful

What they found challenges the conventional wisdom that checking work emails during family time is uniformly harmful. Yes, such “permeability” leads to what the researchers term “ICT fatigue”—a exhaustion born of constant connectivity. This fatigue, in turn, causes people to avoid technology altogether and creates friction with their partners. Conservation of Resources theory, which underpins the study, suggests that humans have finite cognitive and emotional reserves; when these are depleted by round-the-clock availability, something must give.

Yet the same behaviour that drains family resources appears to replenish work ones. Employees who used technology for work purposes during family time reported greater progress towards their professional goals. This progress correlated with higher work engagement and job satisfaction. The explanation, the researchers suggest, lies in autonomy and control. The flexibility to address work matters when it suits, even if that happens to be during nominal family time, allows people to optimize their productivity.

For dual-earner couples, who now represent the majority of American households with children, these trade-offs are particularly acute. Both partners face competing demands from two careers and shared domestic responsibilities. The coordination required, such as who manages the morning video call whilst the other supervises breakfast, whose deadline takes precedence when childcare falls through, makes them a crucible for understanding modern work-life dynamics.

Remote work policies

The findings have implications for how companies approach remote work policies. More than a quarter of professional jobs in America will be remote by the end of this year, and the trend shows no signs of reversing. Yet many firms treat after-hours technology use as either wholly acceptable or entirely forbidden. The reality, this research suggests, is more nuanced.

Organizations might do better to help employees manage permeability rather than eliminate it. That could mean clearer guidelines about when immediate responses are genuinely necessary, or training managers to respect boundaries whilst acknowledging that different people draw those boundaries differently. Some employees may prefer to work intensively during traditional hours and disconnect completely afterwards; others might spread their work across the day, interleaving professional and personal tasks.

There remains, of course, much to understand. The research does not explore how these dynamics vary across different types of work, nor does it examine whether the trade-offs differ for parents versus childless couples.

What is clear, however, is that the genie of remote work will not return to its bottle. The challenge is not to restore some mythical separation between work and home. After all, that ship sailed with the smartphone. Rather, it is to help people navigate permeability in ways that preserve both productivity and peace of mind. As the researchers note, we cannot avoid technology use during family time any more than we can avoid family demands during work hours. The question is how to manage the constant presence of both without being consumed by either.

For now, millions of dual-earner couples are conducting their own experiments, one Zoom call and dinner interruption at a time. The screens that connect them to colleagues also distance them from partners. The tools that boost their careers also strain their relationships. Finding balance in this doubled world may be the defining challenge of modern professional life.

The exodus comes into focus in new research led by Cornell University using a dataset called MIGRATE, which tracks annual moves between American neighborhoods at a resolution roughly 4,600 times finer than standard public sources. Unlike county-level migration figures, typically released every five years, MIGRATE shows that most people displaced by the Camp Fire stayed within Butte County, relocating locally rather than leaving altogether. Such moves are largely invisible in conventional data.

On the move

This matters because America is a nation on the move: roughly 44m people change residence each year. But the bulk of this churn—between neighborhoods rather than counties—escapes official measurement. MIGRATE aims to fill that gap, offering neighbourhood-level migration flows from 2010 to 2019. Its creators argue that a clearer picture of who moves where, and why, could sharpen policymaking on issues ranging from disaster recovery and school funding to housing affordability and segregation.

Existing migration data each has flaws. Government sources such as the census and the American Community Survey are reliable but infrequent, costly, and geographically coarse. Commercial alternatives, which scrape voting records, property transactions, or mobile-phone data, offer greater detail but introduce biases, often overrepresenting wealthier or more digitally connected populations.

MIGRATE attempts to split the difference. Using probabilistic models borrowed from computer science, the researchers combine public records with purchased consumer data to estimate annual migration flows between 47bn pairs of census block groups—small areas that approximate neighborhoods. The result is a granular, if modelled, map of residential mobility.

Clear findings

The early findings are suggestive. Nationally, Americans have become more likely to move into high-income neighborhoods—but not equally so. Residents of predominantly Asian neighborhoods were almost twice as likely as others to move into the richest areas. Those from predominantly Black neighborhoods were twice as likely to end up in the poorest ones. Such disparities complicate narratives about opportunity and underline the persistence of economic sorting.

City-level analysis tells a similar story. In New York, people leaving midtown and lower Manhattan—where incomes are highest—tended to move into other wealthy neighborhoods. Movers from upper Manhattan, by contrast, more often landed in poorer ones. Public-housing residents were less mobile overall, though those who did move showed some signs of upward mobility.

Disasters provide perhaps the clearest test of MIGRATE’s value. After the Camp Fire, residents inside the burn zone were at least three times more likely to move than those just outside it or elsewhere in the county. Comparable patterns followed the 2017 Tubbs Fire in Napa and Sonoma counties. Yet county-level data suggested little change at all.

Clear consequences

Such blind spots have consequences. If policymakers cannot see where people actually go, relief funds and rebuilding efforts risk missing their mark. Finer-grained migration data, the researchers argue, could help ensure that resources follow reality rather than averages.

Whether MIGRATE becomes a staple of social science remains to be seen. But it makes one thing clear: much of America’s most consequential movement happens below the level at which the state usually bothers to look.

Is that what really happens, though? Recent research from Carnegie Mellon set out to explore how the act of turning work into a game can impact workers’ motivation, engagement, and moral agency.

Challenging the norm

The rise of gamification has meant that it’s almost accepted that turning our work into a game can boost our morale and productivity. The researchers wanted to question this heuristic, and particularly to explore what we lose when we do things in order to gain a badge or climb a leaderboard.

The researchers argue that the rewards we get can provide a short-term boost to our motivation and productivity, but they undermine the ethical substance of our professional lives.

Firms increasingly turn to gamification as a bloodless means of discipline, seeking to harness employee enthusiasm without issuing explicit commands. Badges, points, and leaderboards promise alignment between individual effort and institutional goals. Yet this alignment comes at a cost. Motivation is subtly rerouted: tasks are no longer undertaken because they are part of one’s role or values, but because they yield symbolic rewards.

Moral slippage

The research identifies a particular moral slippage in this process. Outcomes may remain commendable, but the character of the action deteriorates. As the paper argues, when systems persistently redirect workers from acting on their own judgments towards chasing instrumental incentives, the moral worth of the work is diminished. Higher productivity, in this view, is no guarantee of ethical improvement; efficiency and virtue are not synonymous.

Case studies underline the point. Former U.S. Air Force drone operators have described missions as resembling video games, cultivating what some call a “PlayStation mentality”. Leaderboards and performance metrics encourage a focus on progression and status rather than the gravity of lethal force. The humanitarian framing of saving lives gives way to the thrill of “levelling up”.

The same logic appears in more prosaic settings. At Omnicare, a pharmaceutical company, a help desk deployed a gamified system that rewarded staff with badges and ranks. One employee spoke with relish of nightly progress and newly earned distinctions. Engagement was high. But the enthusiasm also hinted at a shift towards performance for its own sake, loosening the connection between the work and its ethical purpose.

The broader warning is institutional. When justification is replaced by metrics, a deeper harm emerges: the erosion of a shared moral vocabulary within professions. As measurement crowds out judgment, collective ethical understanding fragments. Gamification, the paper concludes, should therefore be judged not merely by its impact on output, but by its effects on the conditions under which moral agency can still operate.

Public debate has tended to fixate on cheating and assessment. That misses the point. For many students, generative AI has become a routine part of intellectual life: a search engine, explainer, summariser, and sounding board rolled into one. Much of this activity takes place beyond the sight, and sometimes against the rules, of instructors.

A key part of student life

Students increasingly use AI to look up information, clarify concepts, condense reading lists, and test ideas, often without consulting teachers at all. Some have begun to question which academic skills still matter when large portions of intellectual labour can be outsourced to machines. At the same time, they complain that universities offer little guidance on how to use these tools responsibly in ways that match the expectations of employers and society.

This creates an uncomfortable challenge for higher education. Universities are built on assumptions about what it means to learn, to know, and to demonstrate competence. When both knowledge production and problem-solving take place in collaboration with AI, those assumptions need revisiting. The uncomfortable question, as the researchers put it, is what universities offer that machines cannot—and how that value can be made tangible.

Interviews with students and teachers suggest that AI is also becoming a tool for triage. Tasks judged dull, repetitive or irrelevant to future careers are readily handed over to algorithms. Work seen as meaningful or personally formative is prioritised. Teachers worry that this selective delegation may hollow out foundational skills and reduce the practice needed for mastery.

Sluggish response

The institutional response has been sluggish. Universities are designed for stability; technology moves at speed. As a result, digital policy is often reactive: practices change first, rules follow later. The researchers argue that this approach is misguided. Guidelines tied to specific tools quickly become obsolete. More durable would be principle-based rules grounded in academic integrity, transparency, responsibility, and clearly defined learning outcomes.

Such questions cannot be left to individual lecturers to resolve ad hoc. They demand collective decisions about educational goals and standards. The research does not prescribe how AI should be used in higher education. But it shows that students and teachers are already using it in ways that strain established norms—making debates about pedagogy and responsibility increasingly hard to postpone.

The study itself focused on engineering education, drawing on interviews with 25 students who regularly use generative AI. Many described the technology as a tireless, always-available mentor—faster and more responsive than traditional digital tools. Mastery of AI, they believe, is becoming a prerequisite for working life, even as universities lag in providing structured support.

A changing role

Teachers, meanwhile, face a shifting role. When students turn to AI for straightforward questions, interaction with instructors diminishes. Disagreements over what constitutes acceptable use further complicate matters. And under the pressure of heavy workloads, the line between efficiency and cheating becomes harder to police.

A second phase of the research widened the lens. Students from 13 engineering programmes were interviewed individually and in groups, before their views were distilled into five themes. These were then used by teachers and educational developers in a scenario-planning exercise to explore possible futures over the next two years. The aim was not prediction, but preparation.

Whether universities are ready or not, AI is already embedded in student practice. The risk is not that machines will replace higher education—but that higher education will fail to explain why it still matters.

When Spotify purged 75m AI-generated songs from its platform last year, it cited concerns about quality and copyright. The streaming giant was hardly alone in its unease. From Deezer to YouTube, music platforms have wrestled with an uncomfortable question: what happens when machines start making melodies? The conventional wisdom holds that artificial intelligence will supercharge human creativity, helping composers overcome writer’s block and explore new sonic territories. A new study suggests the reality may be rather different.

Researchers at Carnegie Mellon University asked 140 musically trained people to compose a 15-second piano melody. Half were given access to Udio, a text-to-music AI system that conjures complete songs from written prompts. The other half was composed without digital assistance. External listeners then rated the resulting melodies on creativity, musicality, and enjoyment. The findings were striking: compositions influenced by AI scored lower on all three measures. On a five-point scale, AI-assisted melodies scored 0.14 points lower for creativity, which represents a modest but statistically significant decline.

Beyond subjectivity

The researchers attempted to go beyond mere subjective ratings and also analyzed the objective merit of each piece of music. They looked at the composition in its entirety and measured things like the rhythmic complexity, the pitch range, and the note density.

The analysis revealed a clear pattern, with the tunes produced by AI  consistently containing fewer notes, spanning a narrower range of pitches, and generally moving more slowly. They were, in the parlance of music theory, sparser and less kinetic. These structural shifts matter because they correlate strongly with how listeners judge creativity. Fast, dense melodies received higher creativity scores; slower, simpler ones did not.

The researchers propose a couple of explanations for this. The first is that previous studies have strongly linked faster and busier music with heightened states of arousal. We tend to be more alert and engaged, which we associate with novelty and creativity. By contrast, AI errs more towards more relaxed and spacious tunes, which could result in music that is less arousing to our senses.

The second factor is what the researchers refer to as “creative fixation.” Psychological research has long shown that exposure to examples can anchor people’s thinking, narrowing the range of ideas they explore. Generative AI amplifies this effect. Large language models are trained to favour common patterns; unless instructed otherwise, they tend toward the centre of their training distribution. When Udio produces a polished audio track in response to a prompt, that track becomes a powerful reference point, thus shaping the composer’s sense of tempo, density, and melodic contour.

Minimal change

Interestingly, the composers themselves were largely okay with the situation. Those who had used AI said they were as satisfied with the tunes they produced as those without AI support. Their enjoyment levels didn’t drop at all.

AI may make the creative process feel productive and satisfying without actually improving, and perhaps slightly worsening, the quality of what gets produced. For individual creators, this presents a trade-off between the convenience of AI-assisted ideation and the creativity of the final product.

As with all studies, we should be cautious about reading too much into things. The melodies created were generally quite short and simple. They were also produced under controlled conditions that were unlike the complex and multi-layered productions typical of professional music.

The long game

The research also treats creation as a one-time affair, which is obviously not how things work in reality. For instance, while the AI creations were typically seen as less creative and lower quality, it’s also possible that AI will enable composers to generate a lot of ideas very quickly. If the average creativity falls but the number of ideas increases the chances of something exceptional being produced, this might be an acceptable trade-off.

The researchers believe that their findings have broad implications, as the homogenization effect that they found is also being seen in various other creative domains, whether creative writing or visual art. Wherever AI is deployed, we see average creative novelty decline, even if some outliers may buck the trend.

The issue has policy implications as well. Many text-to-music systems are trained on copyrighted recordings without permission. Their creators justify this by invoking fair use, which permits the use of protected material if it generates significant innovation. The thing is, if AI actually lowers creativity, then the argument behind innovation is significantly weaker, and it’s harder to justify running roughshod over intellectual property in order to churn out the equivalent of AI slop.

This doesn’t mean that AI has no value in the music process. For instance, the researchers highlight that tools that help with symbolic composition, which is akin to musical notation, can easily integrate into workflows. AI tools can also provide a degree of granular control, with the composer ultimately in charge of whether to accept or reject suggestions.

For now, the study suggests that caution should be taken before we rush headlong into using AI in creative fields. It’s a technology that can produce benefits, but it can also diminish the very creativity we’re seeking to harness. The challenge for system designers is to harness AI’s generative power while preserving the human capacity for surprise. Otherwise, the future of music risks sounding rather like its past, only less interesting.

Important portrayals

Such portrayals matter because, in everyday life, people increasingly encounter AI not just as software but as something resembling a social actor. Through routine exchanges—receiving assistance, advice, or emotionally attuned responses—users may come to treat AI as part of the social fabric. Yet technical competence alone does not explain why some systems are accepted as companions or partners while others remain tools. The question of how humans and AI might coexist harmoniously is therefore as much social as it is technical.

The study analyses fictional AI characters as relational beings, classifying them along a spectrum of autonomy and social function: from obedient instruments to supportive aides, collaborative peers, and, at the far end, autonomous agents pursuing their own ends. Fiction, the authors argue, serves as a kind of laboratory for moral and ethical experimentation, allowing societies to rehearse futures before they arrive.

On this basis, the researchers propose a model that treats AI not simply as a bundle of functions but as an entity defined by its relationships with humans. Seen as cultural prototypes, fictional AIs illuminate how intelligent machines are framed as social and moral presences. That perspective, they suggest, has implications for the design of interactive and educational systems—and for the broader debate about AI ethics and public acceptance.

Excessive safety

Between 1970 and 1999, courts in several American states adopted “wrongful discharge” exceptions to the country’s employment-at-will doctrine, which traditionally allowed bosses to fire workers “for good cause, bad cause, or no cause at all”. The most sweeping of these reforms, the “good-faith exception”, made it far harder to sack employees without justification. David Wehrheim of IESE Business School in Barcelona and two colleagues at Universidad Carlos III de Madrid seized on this natural experiment to examine how job security shapes the behaviour of corporate researchers.

Their findings challenge the conventional wisdom that employment protection fosters innovation. Tracking over 80,000 inventors across more than 1m patents, the authors found that greater job security did indeed change what inventors produced—but not necessarily for the better. Protected inventors became more likely to draw on their employers’ existing technologies rather than seeking inspiration from outside. Their innovations grew less novel and less disruptive, even as they became more valuable to their current firms.

The mechanism appears to be a straightforward career calculation. When dismissal becomes difficult, an inventor’s expected tenure with their employer lengthens. This makes firm-specific knowledge more attractive to accumulate: mastering the company’s particular technological quirks pays dividends over a longer horizon. But this inward focus comes at a cost. Inventors become less exposed to diverse ideas from beyond the corporate walls—the very cross-pollination that tends to produce breakthrough innovations.

A clear pattern

The pattern held across multiple measures of novelty. Patents filed after employment protection was strengthened scored lower on “disruptiveness”, a metric capturing whether an innovation renders previous knowledge obsolete or merely extends it. Text analysis revealed fewer novel word combinations in patent descriptions. The likelihood of producing a “breakthrough” patent—one in the top 5% by citations received—fell by more than a fifth.

This presents firms with an uncomfortable trade-off. The more derivative innovations that protected inventors produced were indeed more valuable to their current employers, as measured by stock-market reactions to patent grants. But society’s interest in radical innovation may not align with individual firms’ preferences for incremental improvements to existing technologies.

The research carries a sting in the tail for policymakers. Employment protection is often championed as a way to encourage risk-taking by removing the fear of failure. Yet for professional researchers already tasked with innovation, security appears to have the opposite effect. Rather than emboldening inventors to pursue moonshots, it encourages them to mine familiar seams.

Unintended consequences

Several tests bolster this interpretation. The effects were concentrated among younger inventors with stronger career concerns, and among those working far from corporate headquarters, where they enjoyed more autonomy. Critically, what mattered was protection in the inventor’s own state, not the firm’s home state—suggesting individual behaviour, not corporate policy, drove the results.

The findings illuminate a broader tension in innovation management. Organisations naturally tilt toward exploiting their existing knowledge rather than exploring new frontiers. Left to their own devices, inventors might maintain a healthier balance between firm-specific and general learning. But strengthen their job security, and the equilibrium shifts decisively toward exploitation.

This need not spell doom for employment protection. But it does suggest that firms relying heavily on breakthrough innovation may need to design countervailing incentives—perhaps by rewarding inventors for external collaboration or rotating them through diverse projects. The alternative is a workforce that becomes ever more expert in yesterday’s technologies, at precisely the moment when tomorrow’s breakthroughs are needed most.

Economists and workers have fretted about machines stealing jobs since Aristotle pondered self-weaving shuttles in 350 BCE. The Luddites smashed textile looms in 1811; John Maynard Keynes warned of “technological unemployment” in 1930. Today’s panic centers on artificial intelligence. Yet systematic evidence on how technology shapes labor demand over the long run has been scarce—until now.

Changing exposure

A new paper from the Kellogg School of Management represents a heroic effort to measure technology’s impact on work across nearly two centuries. The researchers harness large language models to analyse millions of patent documents dating back to the mid-19th century, comparing their semantic content with descriptions of occupational tasks in each decade.

This allows them to track which jobs faced technological disruption from the automatic telephone exchange (which displaced operators) to the cotton picker spindle (which substituted for textile weavers) to modern computing systems.

Their findings paint a remarkably consistent picture. From 1910 to 2020, technological innovation systematically increased demand for occupations requiring higher education, paying higher wages, and employing more women. A one-standard-deviation rise in an occupation’s exposure to new technology reduced its employment share by around 11-13 percentage points over a decade. But this simple story masks two offsetting forces.

When technology targets all of an occupation’s tasks, like automatic telephone switching for operators, employment collapses. When it automates only specific tasks, say, expense reporting for economists, workers can reallocate their time to other duties, potentially increasing productivity and labor demand.

A key change

The authors’ model captures this through two measures: the average exposure of an occupation’s tasks to technology (which predicts employment declines) and the concentration of that exposure in specific tasks (which partially offsets the damage).

The data fit the theory well. Technology explains roughly one-third of the employment shift toward educated occupations over the past century. It accounts for about half the “hollowing out” of middle-skill jobs, production workers, construction workers, and drivers, relative to both low-skill service workers and high-skill professionals. And it explains nearly half the relative employment growth in female-dominated occupations.

These patterns emerge not merely from recent decades but stretch back to 1910, earlier than many economists had recognised. The conventional wisdom holds that job polarisation began in the 1980s with the rise of computing. The authors find it started much earlier, possibly as far back as the early 20th century. Technology has been pulling women into the labour force and pushing middle-skilled men out for over a century.

A dramatic shift

Yet the direction of innovation has shifted dramatically. Before 1980, patents overwhelmingly related to manual, physical tasks: woodworking machinery, railways, and textile production. The innovations of the late 20th and early 21st centuries increasingly targeted cognitive work. Exposure of manual tasks to technology has consistently predicted job losses throughout the entire period.

But exposure of cognitive tasks tells a time-varying tale: before the 1960s, it correlated with employment gains, suggesting technology complemented thinking work. After the information-and-communications-technology revolution, the relationship turned negative. Recent advances in computing and software now substitute for cognitive labour rather than augmenting it.

This historical shift holds ominous implications for the future. The authors use their model to simulate artificial intelligence’s impact over the next decade. Unlike the technologies of the 20th century, which automated physical labor and complemented cognitive skills, AI threatens to substitute for routine cognitive tasks that do not require substantial experience.

The model predicts a stark reversal: relative labor demand will shift toward occupations with lower education requirements (down 0.65 percentage points annually), lower pay, and more male workers. Managers, clerical staff, and technicians face the steepest declines relative to middle-wage workers.

Some caveats

The projections come with caveats. AI might create new tasks entirely, which the model cannot anticipate. Some applications may complement rather than replace human workers. The analysis focuses on occupational demand, not individual worker outcomes; even if demand for an occupation rises, older workers lacking new skills may struggle. And the predictions concern labour income, not capital income; to the extent that AI profits concentrate among asset-holders, overall inequality could still rise.

Still, the findings suggest that AI may partially reverse two centuries of labor-market evolution. The 20th century saw technology systematically favour the educated and skilled while displacing manual workers. The 21st century’s defining technology appears poised to do the opposite.

Whether this proves a feature or a bug depends on one’s perspective—and one’s education level. For policymakers accustomed to worrying about the “left behind” manufacturing workers, the notion that lawyers and accountants might face similar disruption represents a profound shift. Aristotle’s self-weaving shuttle may finally be coming for the scribes.