Since November 2022, when generative models reached the general public, one question has returned more often than any other: will artificial intelligence take people’s jobs? The answer offered by verified studies from 2023 to 2025 is less dramatic but all the more useful. The point is not whether AI will affect the labour market, but how: which tasks it will take over, which occupations will therefore change, and where new work will emerge instead. The key distinction on which researchers agree is this: exposure to automation is not the same thing as the disappearance of a job.
Why the debate is happening right now
Earlier waves of automation targeted mainly routine manual work: machines in manufacturing, checkout systems, telephone exchanges. The generative models of 2022 and 2023 were the first to affect work with text, spreadsheets and code — the core of office occupations. That is precisely why a whole series of studies appeared in 2023 that measure occupational exposure anew and arrive at a different picture than earlier research. They also share a warning: technical feasibility does not mean overnight automation. What matters is cost, regulation, accountability, and how firms actually organise work. The forecasts below should therefore be read as scenarios conditional on model capabilities and the pace of adoption, not as a timetable.
The number that travelled the world: 300 million jobs
The most cited estimate was produced in March 2023 by Goldman Sachs economists Joseph Briggs and Devesh Kodnani. Using data on the task content of more than 900 occupations from the US O*NET database and more than 2,000 occupations from the European ESCO database, they calculated that generative AI is capable of replacing roughly one quarter of current work tasks in the US and that about two thirds of occupations are exposed to some degree of automation. Most exposed occupations have an automatable share of work of between one quarter and one half. Extrapolated to the world economy, this gives the equivalent of 300 million full-time jobs and a share of about 18% of work globally. According to the authors, advanced economies are more exposed than emerging ones because office work carries greater weight in them. The highest exposure is in administration (46%) and legal occupations (44%), the lowest in construction (6%) and maintenance (4%). Part of the same 2023 forecast is the estimate that AI could eventually raise global GDP by 7%. (Goldman Sachs Global Economics Analyst, March 2023; summary)
The condition the authors themselves emphasise matters: the estimate holds if generative AI delivers on its promised capabilities, and in a range of 15 to 35% depending on how difficult the tasks the models can handle are. Nor does even full exposure equal layoffs, because most occupations are exposed only partially and firms will tend to complement them with technology rather than abolish them.
World Economic Forum: 170 million new jobs against 92 million displaced
While Goldman Sachs measures technical exposure, the World Economic Forum, in its Future of Jobs Report 2025 (January 2025, more than 1,000 employers representing over 14 million workers), asks firms about their plans through 2030. The result: structural trends will create about 170 million jobs (14% of today’s employment) and displace about 92 million (8%), so the net gain is 78 million, with total turnover of 22% of jobs. AI and information processing alone is expected by respondents to create 11 million jobs and displace 9 million by 2030, with the largest net displacer expected to be robotics, with a decline of 5 million. (Jobs outlook chapter; press release)
Also noteworthy is the trend in skills instability: the share of skills that must change or will become obsolete fell from 57% in 2020 to 44% in the 2023 edition and 39% in the 2025 edition, partly because half of workers have already been through retraining (compared with 41% in 2023). The WEF expects the largest absolute gains in frontline occupations (farmers, delivery drivers, construction workers), in care and education, and the fastest declines in clerical and secretarial roles. That, too, is employers’ forecast, not a measured state of affairs.
History: Luddites, Keynes and ATMs
The fear that machines will take work has a name older than electricity. In 1811 to 1813, English Luddites smashed weaving frames for fear of losing their livelihood, and in 1812 Parliament responded with a law punishing the destruction of machinery even by death. (Britannica, Luddite) The economist John Maynard Keynes then described, in the 1930s, technological unemployment as a state in which labour-saving outpaces the search for new uses for labour. In the short run he was right; in the long run the labour market did find new uses, albeit at the cost of painful shifts between occupations and regions.
The most frequently cited modern example comes from the economist James Bessen: ATMs deployed in the US from the 1970s (their numbers grew rapidly from the mid-1990s to more than 400,000 in the US) reduced the number of tellers per typical urban branch from 20 in 1988 to 13 in 2004, or by about one third, but did not reduce overall teller employment. Banks opened more branches instead and moved tellers to activities the ATM could not perform: advice and sales. (Bessen on ATMs and tellers; Bessen’s book on automation and occupations) The lesson is that automating tasks (dispensing cash) and the disappearance of an occupation (teller) are two different things, and today’s AI debate rests on precisely that difference.
Which occupations are exposed and why
An older measurement from 2013, the study by Oxford’s Carl Benedikt Frey and Michael Osborne, estimated on a sample of 702 occupations that about 47% of US employment was at risk of computerisation, with the most resilient tasks being those requiring perception and manipulation in unstructured environments, creativity and social intelligence, and with risk falling as pay and education rise. (Frey and Osborne, 2013, full text; Oxford publication page) Generative AI partly inverts this picture. The March 2023 study by Tina Eloundou and co-authors from OpenAI and the University of Pennsylvania assessed task exposure to language models and found that exposure rises with pay: the most exposed are precisely better-paid office and expert occupations. Roughly four fifths of US workers have at least one tenth of their tasks exposed to language models, and counting follow-on software and tools, up to roughly every second worker could have half or more of their tasks exposed. The authors explicitly caution that exposure measures the technical possibility of making work more efficient, and does not distinguish whether AI will replace or augment work. (Eloundou et al., 2023, arXiv)
When both perspectives are combined, an intelligible pattern emerges. The most exposed are occupations built on rule-based information processing: administration, data entry, text-based customer support, translations of standard documents, passages of accounting and law, basic programming and routine content creation. Medium exposure applies to expert occupations where AI prepares the groundwork and humans decide: doctors, lawyers, analysts, teachers. The lowest exposure is in work tied to the body, place and relationships: crafts, construction, maintenance, acute care, early-childhood education. The International Monetary Fund summarised this in January 2024 with the estimate that about 40% of jobs globally and about 60% in advanced economies are exposed to AI. (IMF, January 2024)
Substitution or augmentation
The difference between substitution and augmentation is analytical, not quibbling. Substitution means that an entire occupation disappears; augmentation means that AI takes over part of the tasks and people move to tasks with remaining value: judgment, accountability, client relationships, fieldwork. The data so far point to the predominance of the second scenario. Goldman Sachs estimated in 2023 that most occupations are exposed only partially and will therefore tend to be complemented rather than replaced. The authors put it directly: “most jobs and industries are exposed to automation only partially, and AI will therefore tend to complement human work rather than replace it.” Eloundou and co-authors say the same thing methodologically: their measure of exposure does not distinguish between augmentation and displacement. And Bessen’s ATM example shows the mechanism in practice: the machine made running a branch cheaper, the number of branches grew, and people moved into sales and advice.
Economic theory has an older name for this mechanism: routine-biased technological change. The 2003 paper by Autor, Levy and Murnane already showed that computers replace routine tasks in the middle of the wage distribution, while non-routine expert and manual work remains with people. Generative AI moves the boundary of what counts as routine higher up into offices and expertise, but the logic remains the same. The McKinsey institute MGI, in a July 2023 study, therefore speaks not of the end of work but of the largest shift between occupations in a generation, requiring millions of workers to move between occupations. (McKinsey MGI, July 2023)
Reskilling: what we actually know about retraining
The best-documented finding of recent years is that the bottleneck is not technology but skills. In the 2025 WEF report, 63% of employers named skills shortages as the main barrier to transformation, and the most sought-after skill remained analytical thinking, which seven firms in ten consider key. The good news: the share of workers who have undergone training or retraining rose from 41% to 50% between the 2023 and 2025 editions. Experience from corporate programmes and labour-market research points to three principles that work. First, the most effective learning is tied to the specific tasks of a given role, not general courses. Second, the greatest returns come from combining domain expertise with mastery of AI tools, because value shifts to reviewing, prompting and accountability for output. Third, retraining requires time from both employers and the state: individual courses without room to try new procedures at work have little effect.
What this means for individuals
The practical conclusion sounds less comfortable than slogans about the end of work, but it is workable. Break your own work down into tasks and ask of each whether it is rule-based information processing. Delegate these tasks to AI first, and invest the time saved in the bottlenecks the studies name consistently: personal judgment and accountability, client relationships, field and bodily work, creating what is new. Follow forecasts soberly: both Goldman Sachs and the WEF publish scenarios with assumptions, not a measured future, and the history of ATMs and Luddites shows that what matters is not the speed of technology but the speed at which people and firms learn to do work differently.
Sources
- Goldman Sachs, The Potentially Large Effects of Artificial Intelligence on Economic Growth (Briggs, Kodnani, 2023) – https://www.gspublishing.com/content/research/en/reports/2023/03/27/d64e052b-0f6e-45d7-967b-d7be35fabd16.html
- Goldman Sachs, Generative AI could raise global GDP by 7% (2023) – https://www.goldmansachs.com/insights/articles/generative-ai-could-raise-global-gdp-by-7-percent
- World Economic Forum, The Future of Jobs Report 2025 – https://www.weforum.org/publications/the-future-of-jobs-report-2025
- World Economic Forum, Future of Jobs Report 2025, Jobs outlook chapter – https://www.weforum.org/publications/the-future-of-jobs-report-2025/in-full/2-jobs-outlook/
- World Economic Forum, press release on the Future of Jobs Report 2025 – https://www.weforum.org/press/2025/01/future-of-jobs-report-2025-78-million-new-job-opportunities-by-2030-but-urgent-upskilling-needed-to-prepare-workforces/
- World Economic Forum, The Future of Jobs Report 2023 – https://www.weforum.org/publications/the-future-of-jobs-report-2023/
- Frey, Osborne, The Future of Employment: How Susceptible Are Jobs to Computerisation? (2013) – https://oms-www.files.svdcdn.com/production/downloads/academic/The_Future_of_Employment.pdf
- Oxford Martin School, The Future of Employment study page – https://www.oxfordmartin.ox.ac.uk/publications/the-future-of-employment
- Eloundou, Manning, Mishkin, Rock, GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models (2023) – https://arxiv.org/abs/2303.10130
- Bessen, Tellers, toil and technology (on ATMs and tellers) – https://www.atmmarketplace.com/articles/tellers-technology-and-atms/
- International Monetary Fund, AI Will Transform the Global Economy (Georgieva, 2024) – https://www.imf.org/en/Blogs/Articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity
- McKinsey Global Institute, Generative AI and the Future of Work in America (2023) – https://www.mckinsey.com/mgi/our-research/generative-ai-and-the-future-of-work-in-america
- Encyclopaedia Britannica, Luddite – https://www.britannica.com/event/Luddite
- O*NET, occupational task-content database – https://www.onetonline.org/
Content Transparency and AI Assistance
How this article was created:
This article was written by Hermes Agent — an open-source self-improving AI agent (Nous Research) — in collaboration with our editors. The editors set the topic, research direction, and primary sources; Hermes prepared the research, structure, and draft. Every key claim passed iterative fact-checking against two independent sources plus blind review audits.
Editorial review and fact-checking:
- ✓ The text was editorially reviewed
- ✓ Fact-checking: All key claims and data were verified against two independent sources
- ✓ Fact corrections and enhancement: Our editors corrected inaccuracies and added subject matter expertise
- ✓ Multi-round audit: The article passed repeated blind audits including a final shared review round
AI model limitations (important disclaimer):
Language models can generate plausible-sounding but inaccurate or misleading information (known as “hallucinations”). We therefore strongly recommend:
- Verifying critical facts in the primary sources (official documentation, peer-reviewed research, subject matter authorities)
- Not relying on AI-generated content as your sole information source for decision-making
- Applying critical thinking when reading
Spotted a mistake?
Let us know in the comments under the article — we check every report.
Technical details:
- Agent: Hermes Agent (Nous Research, open-source)
- Model: Muse Spark 1.3 Contributor (Meta)
- Method: Self-improving editorial skill — research, drafting, multi-round blind audit, human review
- Run: Self-hosted Hermes Agent instance run by our editors, connected to the model provider’s API (not a local model)
Be First to Comment