Job Displacement

The claim is that generative systems perform tasks that firms previously paid people to do: drafting documents, writing and reviewing code, answering customer queries, summarising research, producing routine images and translations. Because those tasks sit inside jobs rather than replacing whole occupations, the first effect is fewer new hires rather than mass firing. Over time, on this account, reduced hiring compounds into a smaller workforce in exposed occupations, and the people affected do not move smoothly into other work.

Two quantities are often confused. Exposure measures the share of an occupation's tasks that a system could perform. Displacement measures jobs that actually disappear. A 2024 International Monetary Fund staff discussion note estimated that around 40% of global employment is exposed to AI, rising to roughly 60% in advanced economies.^[3] A 2025 International Labour Organization working paper by Pawel Gmyrek, Janine Berg and colleagues put about one in four workers worldwide in an occupation with some generative AI exposure, with 3.3% of global employment in the highest exposure gradient, and exposure ranging from 11% of employment in low income countries to 34% in high income countries.^[4] Both bodies stated that transformation of jobs, not elimination, was the more likely outcome for most workers.

The stronger version of the fear rests on speed rather than scale. Earlier automation waves required capital equipment to be designed, financed and installed, which spread the adjustment across decades. Generative systems diffuse through software subscriptions and reach many occupations at once. Forecasts built on that reasoning vary widely. Goldman Sachs economists Joseph Briggs and Devesh Kodnani estimated in March 2023 that generative AI could expose the equivalent of 300 million full time jobs to automation while also noting that new occupations have historically absorbed most displaced workers.^[17] The World Economic Forum's Future of Jobs Report 2025, based on employer surveys, projected 92 million jobs displaced and 170 million created by 2030, a net gain of 78 million.^[18]

The most direct evidence comes from payroll records. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen analysed administrative data from the largest United States payroll processor and reported in November 2025 that workers aged 22 to 25 in the most AI exposed occupations had experienced a relative employment decline of about 16%, while employment for older workers in the same occupations held steady or grew.^[1] Adjustment ran through headcount rather than pay, concentrated where AI substitutes for work rather than assisting it, and persisted when technology firms and remote eligible roles were excluded. A February 2026 update found that under the strictest firm level controls the decline became significant only from 2024, and that interest rates do not explain it.^[2]

Survey evidence tracks worry rather than outcomes. Pew Research Center, surveying 5,273 employed United States adults in October 2024, found 52% worried about future AI use in the workplace against 36% hopeful, with 32% expecting fewer job opportunities for themselves in the long run and 6% expecting more.^[12] Gallup, surveying 23,717 United States employees in February 2026, found 18% saying it was very or somewhat likely their job would be eliminated within five years because of AI or automation, rising to 23% among employees whose organisation had adopted AI.^[11]

Employer statements are a third strand. Amazon announced a reduction of roughly 14,000 corporate roles in October 2025, attributing it to a need to operate with fewer layers alongside the arrival of AI.^[9] Its chief executive had set out the reasoning earlier that year.

We will need fewer people doing some of the jobs that are being done today, and more people doing other types of jobs. It's hard to know exactly where this nets out over time, but in the next few years, we expect that this will reduce our total corporate workforce as we get efficiency gains from using AI extensively across the company.

Against that, the aggregate United States labour market in mid 2026 was slack rather than collapsing. The Bureau of Labor Statistics reported unemployment at 4.2% in June 2026 with nonfarm payrolls up 57,000.^[24] The Federal Reserve Bank of New York recorded unemployment for recent college graduates aged 22 to 27 at 5.6% in early 2026, above the national rate and a reversal of the long standing graduate advantage.^[23] The second International AI Safety Report, chaired by Yoshua Bengio and released on 3 February 2026 by a panel of 96 experts, treats the integration of AI into labour markets as a systemic concern while stopping short of finding a measured effect on overall employment.^[22] The Stanford HAI AI Index reached the same reading from the other direction, recording employer expectations of cuts alongside the observation that large scale job losses have not yet shown up in overall employment data.^[13]

The most specific mechanism in the current evidence is the loss of the entry rung. Junior roles in software, customer support and analysis existed partly because the work needed doing and partly because they were how people became senior. If firms buy the first function from a model and keep only experienced staff, the supply of experienced staff falls a decade later. The Stanford HAI AI Index recorded employer expectations of reductions concentrated in service operations, supply chain and software engineering, with anticipated cuts exceeding those already observed.^[13]

A second mechanism is that adjustment is local and slow even when the aggregate numbers look fine. The OECD Employment Outlook 2026 found that regional labour markets exposed to trade and technology shocks tend to adjust through transitions into joblessness and through different hiring for newcomers rather than through displaced workers moving between sectors, leaving lasting earnings damage.^[25] National employment can hold up while particular towns and cohorts do not recover.

A third is measurement failure. New York added a disclosure field to its WARN mass layoff notices in March 2025, asking employers whether technological innovation or automation contributed, and in the first year no employer attributed a layoff to AI, despite more than 160 companies filing notices.^[20] Firms that restructure around AI typically describe the result as reorganisation or cost reduction. If official reporting depends on employers naming the cause, displacement can be real and statistically invisible at the same time.

The fourth is that firms may cut headcount on the expectation of AI gains that do not materialise. Challenger, Gray and Christmas recorded AI as the leading stated reason for announced United States job cuts for four consecutive months in 2026.^[8] Announcements are decisions, not measurements of productivity, and the cost of a wrong decision falls on the worker either way.

Daron Acemoglu of MIT has made the strongest quantitative case against large near term displacement. Building the estimate from the fraction of tasks AI can affect and the average cost saving on those tasks, he put the total factor productivity effect at no more than 0.66% over ten years, and argued the figure should be revised down to less than 0.53% once the limits of task level savings are allowed for.^[14] That is roughly 0.05% to 0.07% a year, consequences he called nontrivial but modest. If the productivity gain is that small, the cost saving that would justify substituting capital for large numbers of workers is not there either.

The aggregate data has so far sided with the sceptics. Molly Kinder of Brookings with Martha Gimbel, Joshua Kendall and Maddie Lee of the Budget Lab at Yale examined the 33 months after the ChatGPT release and found the share of workers in high, medium and low exposure occupations close to flat, with the occupational mix shifting only marginally faster than during the computer and internet transitions.^[5] They were explicit about the limits of the method, and about the possibility that it misses effects on particular occupations or subpopulations, early career workers among them.

Our methodology might miss the labor market equivalent of a small fire starting on the stove, but would clearly detect if the house was burning down.

Richard Audoly, Miles Guerin and Giorgio Topa of the Federal Reserve Bank of New York examined job postings in May 2026 and reached a similar conclusion by a different route. They found that the relative decline in vacancies for AI exposed occupations had begun before ChatGPT was released, and that junior postings within those occupations did not diverge from senior ones, which is the opposite of what the entry level compression story predicts.^[6]

There is also direct evidence that deployment does not automatically deliver the savings firms assume. METR ran a randomised controlled trial with 16 experienced open source developers across 246 tasks in repositories they knew well, and found that allowing AI tools made them 19% slower, while the same developers believed the tools had sped them up by about 20%.^[16] A gap that large between perceived and measured effect is a reason to treat both employer optimism and displacement projections as provisional.

The historical base rate cuts the same way. David Autor, Caroline Chin, Anna Salomons and Bryan Seegmiller found that around 60% of United States employment in 2018 was in job titles that did not exist in 1940.^[15] That does not guarantee a repeat, and the same paper records the location of new work shifting from middle paid to high paid and low paid roles since 1980, which is itself a distributional problem. It does mean that a count of tasks a model can perform is not a count of jobs that will vanish.

United States federal policy has focused on measurement and retraining rather than restriction. America's AI Action Plan, released in July 2025, directed the Department of Labor to create an AI Workforce Research Hub to track AI adoption, job creation, displacement and wage effects, to prioritise AI skills in existing workforce funding streams, to use discretionary funds for rapid retraining of workers displaced by AI, and to issue guidance to states on identifying dislocated workers in affected sectors.^[19]

States have moved on disclosure. New York became the first to ask employers filing mass layoff notices whether automation or AI contributed, a field added in March 2025, and legislators have since proposed an Artificial Intelligence Workforce Impact Transparency Act to formalise the requirement in statute.^[20] California took a different route and stopped short: the legislature passed Senate Bill 7, the No Robo Bosses Act, which would have required notice before automated decision systems were used in employment decisions and barred employers from relying on them alone for discipline or termination, and Governor Gavin Newsom vetoed it on 13 October 2025, objecting to unfocused notification requirements and overly broad restrictions. The legislature let the veto stand on 2 March 2026.^[26]

Collective bargaining has produced the most concrete job guarantees. The International Longshoremen's Association and the United States Maritime Alliance reached a six year deal in January 2025 covering roughly 20,000 workers on the East and Gulf Coasts, combining a 62% wage increase with permission to introduce semi automated rail mounted gantry cranes in exchange for a commitment to add one job per semi automated crane, including at ports where the machines were already installed.^[21] Labor Notes noted at the time that this is still fewer workers than conventional operations require, and that some of the added posts are maintenance rather than crane operation. The structure of the agreement, technology allowed in return for a headcount commitment per unit of equipment, is the template unions in other sectors have cited.

Monitoring infrastructure has improved faster than policy. The ILO maintains a refined global index of occupational exposure to generative AI, the Stanford Digital Economy Lab has continued to publish updated payroll analyses of employment by AI exposure, and the second International AI Safety Report, released on 3 February 2026, reviews labour market evidence for the thirty countries that commissioned the series.^[4][2][22]

Keep three quantities apart. Exposure is what a model could do. Adoption is what firms have actually deployed into workflows. Displacement is jobs that ended. Most alarming headline numbers are exposure figures, and exposure has never mapped one to one onto job loss. Conversely, the reassuring aggregate statistics measure the whole labour market, where a large effect on a small group is invisible. Both the payroll studies finding entry level declines and the household survey studies finding stability can be correct at once, because they are counting different things.

Pay attention to the data source when a study contradicts another. Administrative payroll records catch changes in specific firms and age groups quickly. Household surveys catch economy wide composition slowly and with sampling noise. Job posting data catches intentions rather than employment. Employer announcements catch narratives, and firms have incentives to describe cost cutting as AI modernisation when investors reward it. When the four disagree, they are usually disagreeing about coverage rather than about facts.

The question that separates a manageable transition from a serious one is not whether total employment falls, but whether people displaced from exposed occupations find comparable work, and whether firms keep hiring at the bottom. Those are measurable, and the series are published.