Automation, AI and jobs: what the evidence shows and what is still forecast

Claims about artificial intelligence and jobs range from "half of all jobs will disappear" to "nothing has changed". Most of these claims mix three different kinds of statement: what has been measured about past automation, what has been estimated about which tasks new technologies could perform, and what is forecast about the future. This study separates them. It reviews the best-known research on industrial robots, the widely cited estimates of jobs "at risk" from computerisation and their critiques, the exposure studies of large language models by researchers at OpenAI and the International Labour Organization, the field and laboratory experiments on generative AI and productivity, and the first studies of actual labour-market outcomes in 2023–2026. The established findings are narrower than public debate suggests: automation has displaced workers in specific places and groups, while aggregate employment has so far adjusted; generative AI raises productivity in some tasks, especially for less experienced workers; and early administrative data show no economy-wide displacement but a growing gap in employment for young workers in the most exposed occupations. Everything beyond that is forecast.

Key findings

• Robots have had measurable local effects. In US commuting zones, one more industrial robot per thousand workers reduced the employment-to-population ratio by 0.2 percentage points and wages by 0.42% (Acemoglu and Restrepo, 2020). In Germany, robot-related job losses in manufacturing were fully offset by new service jobs, with the burden falling mainly on young workers entering the labour force (Dauth et al., 2021). Across 17 countries in 1993–2007, robots raised annual labour productivity growth by about 0.36 percentage points without significantly reducing total employment (Graetz and Michaels, 2018).
• "Jobs at risk" estimates are not forecasts of job loss. Frey and Osborne's 2013 estimate that about 47% of US employment was at high risk of computerisation described technical potential "over some unspecified number of years, perhaps a decade or two". A task-based reanalysis for the OECD found that on average 9% of jobs across 21 OECD countries were automatable.
• Large language models are exposed to many tasks. Eloundou and colleagues estimate that around 80% of the US workforce could have at least 10% of their work tasks affected by large language models, and about 19% at least half; with LLM alone about 15% of tasks could be done significantly faster at the same quality, rising to 47–56% with software built on top.
• The ILO expects transformation more than automation. Its 2025 index finds that one in four workers worldwide is in an occupation with some generative-AI exposure, and 3.3% of global employment is in the highest exposure category — 4.7% of female employment against 2.4% of male. Exposure is 34% of employment in high-income countries against 11% in low-income countries.
• AI assistance raises productivity in measured tasks. Among 5,172 customer-support agents, access to an AI assistant raised issues resolved per hour by 15% on average, with the largest gains for less experienced and lower-skilled workers (Brynjolfsson, Li and Raymond, 2025). In an experiment with 453 professionals, ChatGPT cut the time for writing tasks by 40% and raised quality by 18% (Noy and Zhang, 2023).
• Early outcome data show no aggregate displacement. Danish administrative data show precise null effects of AI chatbots on earnings and hours, ruling out effects larger than 2% two years after ChatGPT's launch (Humlum and Vestergaard, 2025). US payroll data through June 2026 show no widespread displacement, but employment of 22–25-year-olds in AI-exposed occupations stands 19% below where it would be had it kept pace with less-exposed peers — a descriptive gap, not a causal estimate (Brynjolfsson, Chandar and Chen, revised 2026).
• Macroeconomic gains may be modest. Acemoglu (2025) estimates that AI will raise total factor productivity by no more than 0.66% over ten years, and possibly less than 0.53%.

Research question and scope

This study asks what research has established, and what it has not, about the effects of automation and AI on employment, wages and the organisation of work. We are interested in this question because it underlies the "post-labour" debate: whether societies need new institutions — basic incomes, social wealth funds, new forms of property — because demand for human labour will fall.

The scope covers peer-reviewed and prominent working-paper research from 2013 to 2026 in five categories: local labour-market studies of industrial robots; occupational "risk" estimates; exposure indices for generative AI; experiments on productivity; and early administrative evidence on labour-market outcomes since the release of ChatGPT in November 2022. We do not review industry forecasts, consultancy reports or surveys of executives' intentions, which are frequently cited but rarely transparent about method. Other atlas studies cover the evidence on income-support policies that might respond to labour-market change.

Background

An old question

Fears of technological unemployment are as old as industrial economics. John Maynard Keynes coined the phrase "technological unemployment" in 1930, describing it as unemployment due to discovering means of economising the use of labour outrunning the pace at which new uses for labour can be found. Since then, employment in rich countries has grown through repeated waves of mechanisation and computerisation.

Why jobs have survived — so far

David Autor's 2015 essay "Why Are There Still So Many Jobs?" summarises the mainstream explanation. Automation does substitute for labour, as it is intended to, but it also complements labour, raises output in ways that increase demand for labour, and interacts with adjustments in labour supply. Autor argues that commentators tend to overstate machine substitution and ignore these complementarities. Technology does change which jobs are available and what they pay: in recent decades, a "polarisation" of the labour market saw wage gains go disproportionately to those at the top and bottom of the skill distribution, not those in the middle.

Tasks, not jobs

The modern framework, developed by Daron Acemoglu and Pascual Restrepo among others, treats jobs as bundles of tasks. In their 2018 model, automation of existing tasks reduces employment and the labour share and may reduce wages, while the creation of new tasks in which labour has a comparative advantage has the opposite effect. Whether labour's position improves or deteriorates depends on the race between these two forces. In a version with heterogeneous skills, inequality rises during transitions driven by either faster automation or new tasks. This framework explains why "how many jobs does AI threaten?" is the wrong question: what matters is which tasks are automated, which new tasks emerge, and how fast workers can move between them.

Method and sources

Five kinds of evidence

How the atlas grades claims

We treat findings from the first and fourth categories as established where they come from peer-reviewed studies with credible identification, while recognising that they describe particular technologies, places and periods. We treat exposure estimates as established as measurements but not as predictions. We treat early outcome monitoring as emerging: informative, but too recent to be conclusive. We treat any statement about future aggregate employment as a forecast.

Findings: industrial robots

United States

Acemoglu and Restrepo (2020) study the effects of industrial robots on US labour markets, using variation in exposure to robots defined from industry-level advances in robotics and each commuting zone's industry mix. They find robust negative effects on employment and wages: one more robot per thousand workers reduces the employment-to-population ratio by 0.2 percentage points and wages by 0.42%. The areas most exposed to robots after 1990 do not show differential trends before then, and robots' impact is distinct from that of other capital and technologies.

Germany

Dauth, Findeisen, Suedekum and Woessner (2021) use detailed administrative data to study the adjustment of German local labour markets to industrial robots. Robot exposure is associated with job displacement in manufacturing, but those losses are fully offset by new jobs in services. The burden falls mostly on young workers just entering the labour force, who substitute away from vocational training towards colleges and universities. Incumbent workers see more stable employment within their firms, often by taking on new tasks in their original plants, and workers in occupations with complementary tasks, such as managers and technical scientists, benefit.

Cross-country evidence

Graetz and Michaels (2018) analyse robot adoption in 17 countries from 1993 to 2007. Increased robot use contributed approximately 0.36 percentage points to annual labour productivity growth, while raising total factor productivity and lowering output prices. They find that robots did not significantly reduce total employment, although they did reduce low-skilled workers' share of employment.

What the robot studies establish

The three studies are not contradictory. Automation can reduce employment and wages in exposed places and groups while aggregate employment adjusts, depending on institutions, demand and the pace of adoption. The distributional effects — on regions, young entrants and low-skilled workers — are better established than any aggregate effect.

Findings: estimates of jobs "at risk"

Frey and Osborne

In a working paper dated 17 September 2013, later published in Technological Forecasting and Social Change (2017), Carl Benedikt Frey and Michael Osborne estimated the probability of computerisation for 702 detailed occupations, using expert labels for a subset and a statistical classifier for the rest. They concluded that about 47% of total US employment was in the high-risk category, "meaning that associated occupations are potentially automatable over some unspecified number of years, perhaps a decade or two". They also found a strong negative relationship between an occupation's probability of computerisation and its wages and educational attainment.

Critiques

The 47% figure became one of the most quoted numbers in the automation debate, often as a forecast that half of jobs would disappear. Three critiques are important.

1. Occupations are not tasks. Arntz, Gregory and Zierahn (2016) argued for the OECD that assuming whole occupations are automated overstates automatability, because occupations labelled high-risk often still contain a substantial share of tasks that are hard to automate. Taking account of task heterogeneity within occupations, they found that on average 9% of jobs across 21 OECD countries were automatable, ranging from 6% in Korea to 12% in Austria.
2. Technical potential is not adoption. Whether a task is automated depends on cost, regulation, organisational change and demand, not only on technical feasibility. Frey and Osborne explicitly described a "rough timeline" rather than a prediction of job losses.
3. Individual-level estimates differ. Nedelkoska and Quintini (2018) for the OECD built on Frey and Osborne's expert assessment but estimated risk for individual jobs using the Survey of Adult Skills in 32 countries, linking risk to skills use and training. Their approach, like Arntz and colleagues', yields estimates well below the occupation-based figure.

A decade later

More than a decade after the 2013 paper, US employment has not fallen by anything resembling 47%. This does not prove Frey and Osborne wrong, since they did not predict a date, but it illustrates why technical-potential estimates should not be read as forecasts.

Findings: generative AI exposure

"GPTs are GPTs"

Eloundou, Manning, Mishkin and Rock assessed US occupations for their alignment with large language model capabilities, combining human annotators and GPT-4 classifications. Their first version was released in March 2023; the study was published in Science in 2024. They estimate that around 80% of the US workforce could have at least 10% of their work tasks affected by LLMs, and about 19% of workers may see at least 50% of their tasks affected. With access to an LLM, about 15% of all worker tasks could be completed significantly faster at the same quality; with software and tooling built on LLMs, this rises to 47–56%. The projected effects span all wage levels, with higher-income jobs potentially facing greater exposure. The authors explicitly make no predictions about the timeline of development or adoption, and argue that LLMs have the traits of general-purpose technologies.

Two cautions apply. "Affected" means that a task could be performed faster, not that a worker would be replaced. And the classification partly relies on a language model judging its own potential, which raises questions about calibration.

The ILO's 2023 analysis

The ILO's first global study (Gmyrek, Berg and Bescond, August 2023) assessed exposure to GPT-4 and concluded that the overwhelming effect of the technology was likely to be augmentation of occupations rather than automation. Jobs with high automation potential accounted for only 0.4% of employment in low-income countries but 5.5% in high-income countries. The augmentation potential was larger, affecting 10.4% of employment in low-income and 13.4% in high-income countries. The effects were highly gendered: in high-income countries, jobs with high automation potential made up 8.5% of female employment against 3.9% of male employment, largely because clerical jobs are an important source of women's employment.

The ILO's 2025 refined index

The 2025 update (Gmyrek et al.) rebuilt the index using a sample from 29,753 tasks in the Polish occupational classification, a survey of 1,640 workers across all major occupational groups yielding 52,558 data points on 2,861 tasks, and Delphi-style rounds with international experts. It finds that clerical occupations continue to have the highest exposure, that some highly digitised professional and technical occupations have become more exposed, and that globally one in four workers is in an occupation with some generative-AI exposure. 3.3% of global employment falls in the highest exposure gradient — 4.7% of female against 2.4% of male employment, a gap that widens to 9.6% against 3.5% in high-income countries. Overall exposure is 11% of employment in low-income and 34% in high-income countries. The ILO concludes that, because most occupations consist of tasks requiring human input, transformation of jobs is the most likely impact.

Findings: productivity experiments

Customer support

Brynjolfsson, Li and Raymond study the staggered introduction of a generative-AI conversational assistant to 5,172 customer-support agents. Access to AI assistance increased productivity, measured as issues resolved per hour, by 15% on average, with substantial heterogeneity. Less experienced and lower-skilled workers improved both the speed and quality of their output, while the most experienced and highest-skilled workers saw small gains in speed and small declines in quality. AI assistance appeared to help workers learn and improved English fluency, particularly among international agents. Gains were largest for moderately rare problems. Customers were more polite and less likely to ask for a manager.

Professional writing

Noy and Zhang assigned occupation-specific, incentivised writing tasks to 453 college-educated professionals and randomly gave half access to ChatGPT. Average completion time fell by 40% and output quality rose by 18%. Inequality between workers decreased. Workers exposed to ChatGPT in the experiment were twice as likely to report using it in their real job two weeks later and 1.6 times as likely two months later.

What the experiments show and do not show

These studies establish that current generative-AI tools can raise productivity substantially in well-defined tasks and that gains tend to be larger for less experienced workers — narrowing gaps within those tasks. They do not show what happens to employment when such gains spread. Higher productivity per worker can reduce the number of workers needed, raise output and demand, or shift workers to new tasks; the experiments cannot tell which will dominate.

Findings: early labour-market outcomes, 2023–2026

Denmark: rapid adoption, no effect on earnings yet

Humlum and Vestergaard link large-scale surveys of chatbot adoption to Danish administrative labour-market records. They document "rapid currents": most employers in exposed occupations had adopted chatbot initiatives, workers reported productivity benefits, and new AI-related tasks were widespread. Yet using difference-in-differences methods, they estimate precise null effects on earnings and recorded hours at both worker and workplace level, ruling out effects larger than 2% two years after the launch of ChatGPT. What moved was the structure of work: employers absorbed AI by reorganising tasks, including new tasks in content generation, AI oversight and AI integration.

United States: a widening gap for young workers

Brynjolfsson, Chandar and Chen use high-frequency payroll data from ADP covering millions of US workers. Their paper, first released in August 2025 and revised in August 2026 with data through June 2026, finds no evidence of widespread, economy-wide job displacement. However, employment of young workers aged 22–25 in AI-exposed occupations now stands 19% below where it would be had it kept pace with less-exposed peers, a gap that has widened steadily since August 2025; experienced workers show no comparable gap. The divergence operates mainly through reduced hiring of young workers rather than increased separations, and is concentrated in occupations where AI use mainly substitutes for human tasks; where it mainly complements workers, employment is flat or rising. The pattern persists when technology firms and computer occupations are excluded, and when controlling for interest-rate exposure and remote work. It attenuates when controlling for education, shows some divergent trends before generative AI, and is more pronounced in the ADP sample than in national survey benchmarks. The authors describe these facts as early, descriptive indicators — "canaries in the coal mine" — rather than causal estimates.

Macroeconomic expectations

Acemoglu's "The simple macroeconomics of AI" (2025) uses a task-based model and existing estimates of exposure and task-level productivity gains to estimate aggregate effects. He finds macroeconomic effects that are non-trivial but modest — no more than a 0.66% increase in total factor productivity over ten years — and argues that even this may be too high, because early evidence comes from easy-to-learn tasks, giving a revised estimate of less than 0.53%. He also argues that AI is unlikely to increase inequality as much as previous automation technologies, but that there is no evidence it will reduce labour-income inequality, and that it is predicted to widen the gap between capital and labour income.

Separating evidence from forecast

What the evidence does not show

• It does not show that AI will cause mass unemployment. No study reviewed here provides evidence of aggregate job loss from generative AI to date, and exposure estimates are not displacement estimates.
• It does not show that AI will not cause large job losses. Early outcome studies cover two to three years after a technology whose capabilities and adoption are changing rapidly. Past automation took decades to show its full effects.
• It does not show that past automation is a reliable guide. Robot studies concern a technology mainly used in manufacturing. Generative AI affects cognitive and clerical tasks across sectors, including higher-wage jobs.
• It does not show that productivity gains will be broadly shared. Experiments measure output per worker in a task, not who captures the gains.
• It does not identify the cause of the young-worker gap. The payroll evidence is descriptive and partly attenuated by education controls and pre-existing trends.
• It does not settle whether exposure means augmentation or substitution. The ILO and Eloundou indices measure what could be affected; the same exposure can raise or lower demand for labour depending on how firms deploy the technology.

Implications for institution design

1. Build monitoring before building responses. The most useful recent work links adoption surveys to administrative payroll and tax records. Governments and statistical agencies should create public, regularly updated indicators of AI adoption, hiring, earnings and task change by occupation and age, so that policy can respond to evidence rather than to forecasts.
2. Watch entry-level hiring. If AI reduces the hiring of young workers into exposed occupations, the first institutions to strain will be those that train and place new entrants: apprenticeships, graduate schemes and the early-career rungs of professional ladders. Robot studies in Germany also found that the burden fell on young entrants.
3. Prepare income support that can scale with evidence. The cash-transfer evidence reviewed elsewhere in the atlas suggests modest labour-supply effects of unconditional payments. Insurance-like mechanisms that expand automatically when measured displacement rises would be more defensible than either premature universal schemes or no plan at all.
4. Treat the capital–labour split as a design question. If AI widens the gap between capital and labour income, as Acemoglu predicts, debates about social wealth funds, profit-sharing or "AI dividends" become questions about ownership, not only about benefits.
5. Support augmentation, not only displacement. The experiments show that AI tools can raise the productivity of less experienced workers. Training, procurement and workplace consultation can steer deployment towards complementing workers — one of the few levers available before outcomes are known.
6. Label forecasts as forecasts. Public debate would benefit from the discipline used in this study: separating what has been measured from what has been modelled and from what is predicted, each with a date and source.

Open questions

• Will the employment gap for young workers in AI-exposed occupations persist, widen or close, and is AI its main cause?
• How quickly will task reorganisation inside firms — new tasks in AI oversight and integration — translate into changes in wages and occupational structure?
• Will generative AI follow the pattern of earlier automation, with local and distributional effects but stable aggregate employment, or will its breadth across sectors produce larger aggregate effects?
• How will the gendered exposure of clerical work affect women's employment, especially in high-income countries?
• What share of AI-driven productivity gains will accrue to workers, consumers and owners of capital?

Sources

1. Acemoglu, D. and Restrepo, P. (2020). "Robots and Jobs: Evidence from US Labor Markets." Journal of Political Economy 128(6): 2188–2244. doi:10.1086/705716
2. Acemoglu, D. and Restrepo, P. (2018). "The Race between Man and Machine: Implications of Technology for Growth, Factor Shares, and Employment." American Economic Review 108(6): 1488–1542. doi:10.1257/aer.20160696
3. Dauth, W., Findeisen, S., Suedekum, J. and Woessner, N. (2021). "The Adjustment of Labor Markets to Robots." Journal of the European Economic Association 19(6): 3104–3153. doi:10.1093/jeea/jvab012
4. Graetz, G. and Michaels, G. (2018). "Robots at Work." Review of Economics and Statistics 100(5): 753–768. doi:10.1162/rest_a_00754
5. Frey, C. B. and Osborne, M. A. (2013). The Future of Employment: How Susceptible Are Jobs to Computerisation? Oxford Martin School working paper. PDF; published in Technological Forecasting and Social Change 114 (2017): 254–280, doi:10.1016/j.techfore.2016.08.019
6. Arntz, M., Gregory, T. and Zierahn, U. (2016). The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis. OECD Social, Employment and Migration Working Papers 189. doi:10.1787/5jlz9h56dvq7-en
7. Nedelkoska, L. and Quintini, G. (2018). Automation, skills use and training. OECD Social, Employment and Migration Working Papers 202. doi:10.1787/2e2f4eea-en
8. Autor, D. H. (2015). "Why Are There Still So Many Jobs? The History and Future of Workplace Automation." Journal of Economic Perspectives 29(3): 3–30. doi:10.1257/jep.29.3.3
9. Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2024). "GPTs are GPTs: Labor market impact potential of LLMs." Science 384(6702): 1306–1308. doi:10.1126/science.adj0998; working paper arXiv:2303.10130
10. Gmyrek, P., Berg, J. and Bescond, D. (2023). Generative AI and Jobs: A global analysis of potential effects on job quantity and quality. ILO Working Paper 96. doi:10.54394/FHEM8239
11. Gmyrek, P., Berg, J., Kamiński, K. et al. (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper. doi:10.54394/HETP0387
12. Brynjolfsson, E., Li, D. and Raymond, L. (2025). "Generative AI at Work." Quarterly Journal of Economics 140(2): 889–942. doi:10.1093/qje/qjae044
13. Noy, S. and Zhang, W. (2023). "Experimental evidence on the productivity effects of generative artificial intelligence." Science 381(6654): 187–192. doi:10.1126/science.adh2586
14. Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. doi:10.3386/w33777
15. Brynjolfsson, E., Chandar, B. and Chen, R. (2025, revised 2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. digitaleconomy.stanford.edu
16. Acemoglu, D. (2025). "The simple macroeconomics of AI." Economic Policy 40(121): 13–58. doi:10.1093/epolic/eiae042