AI job-loss statistics, checked
Updated · 10 sources
Headline figures about AI and jobs measure different things. Some model possible job losses, some estimate exposure, and others describe changes in a sample of workers. None tells you how employment will change in your town. Check each study's date, country and assumptions before treating a claim as a forecast.
Claims in context
- “Millions of UK jobs will be lost”
- IPPR modelled scenarios. Its worst case assumes every exposed job is replaced; its best case has no job losses. Neither is a prediction, though IPPR says the worst case is a real possibility without government action. Institute for Public Policy Research.
- “Most UK jobs will be replaced”
- The figure behind this is an IMF exposure estimate, quoted by the government, that counts work AI could enhance as well as work it could perform. Department for Science, Innovation and Technology and the AI Security Institute.
- “Almost half of jobs will disappear”
- An older US estimate of technical susceptibility, not observed losses. Oxford Martin School.
- “Microsoft found jobs AI cannot replace”
- A US preprint that ranked where chatbot use applies to occupations; it did not forecast replacement. Microsoft Research (arXiv preprint).
- “ONS proved current AI job losses”
- Its England model estimates task automation and predates widespread generative AI. Office for National Statistics.
- “AI is already behind mass UK job losses”
- Around half of UK businesses using AI report no change in headcount in the ONS survey, and few report cuts; the government says causation is not yet established. Office for National Statistics.
- “The Stanford finding proves a general crash”
- A US payroll sample found a relative gap for young workers, not economy-wide displacement. Stanford Digital Economy Lab.
- “The King's result covers every UK firm”
- The press release describes firms with differing AI exposure; we could not open the working paper itself. King's College London.
What the evidence says
IPPR's scenarios
The widely repeated 7.9 million is IPPR's worst case for a second wave of adoption, in which firms build AI deeply into their work. For that wave its central case is 4.4 million, and its full-augmentation case has zero modelled job losses. For the tools available when IPPR wrote, its worst case was 1.5 million.[1] These are alternative scenarios, not forecasts, though IPPR warns that without government action its worst case is a real possibility.
Source: Institute for Public Policy Research · 27 March 2024 · UK
The UK exposure share
The government's evidence review quotes IMF estimates that around 70 per cent of UK workers are in occupations with tasks AI could perform or enhance. By the same estimates, around half of those workers are in roles where AI is more likely to support people, and the other half in roles where it may be more likely to do their tasks.[2]
Source: Department for Science, Innovation and Technology and the AI Security Institute · 28 January 2026 · UK
An older US estimate
In a US study, long before today's generative-AI tools, Frey and Osborne estimated that about 47 per cent of US employment was technically susceptible to computerisation.[3] They asked which jobs computers could take over, not how many jobs would be lost.
Source: Oxford Martin School · 1 September 2013 · US
Why studies disagree
An ONS model of jobs in England, built before generative AI, gave its highest automation scores to routine, low-skilled work. On a scale from zero to one, the England model scored programmers 0.27 and electricians 0.51.[4] Newer generative-AI studies rank office and professional work as more exposed.
Source: Office for National Statistics · 25 March 2019 · England
Early employment signals
Stanford's revised US payroll study found that employment of workers aged twenty-two to twenty-five in AI-exposed occupations was 19 per cent below where it would have been had it kept pace with less-exposed peers.[5] It found no economy-wide displacement, and its authors call these early, descriptive indicators rather than proof that AI caused the gap. In a press release, King's College London reported that junior positions at highly exposed UK firms fell by 5.8 per cent.[6]
Source: Stanford Digital Economy Lab · 12 August 2026 · US
What UK businesses report
In the ONS business survey, around half of businesses using AI said it had not changed their headcount, and many were unsure.[7] The government's review says that falling hiring in more exposed occupations does not yet show that AI is the cause.
Source: Office for National Statistics · 20 July 2026 · UK
What it doesn’t tell you
- A modelled worst case is one branch of a scenario exercise. It cannot be reported as observed UK job losses.
- The exposure share counts work AI could enhance as well as work it could perform. It is the IMF's estimate as the government quotes it; we could not open the IMF note itself.
- The Oxford estimate used US occupations and predates today's generative-AI tools.
- The ONS model covers England only and was published before generative AI was in wide use.
- The Stanford result is a revised US sample. Its authors call their findings early, descriptive indicators rather than causal estimates.
- The King's press release describes an earlier version of a working paper we could not open. The government's review summarises a later version: postings fell more in more exposed occupations, mainly higher-paid ones, and returned to their original levels after around 20 months.[2]
- Microsoft's study is a US preprint about where chatbot use applies, not a measure of jobs lost.
Check the denominator before sharing a figure
Ask what was counted: tasks, workers, businesses, vacancies or hours. Then check the country, observation period and whether the number is measured or modelled. If a headline changes any of those labels, do not carry its percentage into a career decision. This is a reading checklist, not a new estimate.
Read the source: Department for Science, Innovation and Technology and the AI Security Institute
If you’re weighing up a trade
Treat a viral claim as a question about method, then compare your own tasks and the practical route into a trade before paying for training.
Common questions
- Did IPPR forecast UK job losses?
- No. It modelled alternative scenarios whose outcomes depend on assumptions and policy choices.
- Does the IMF measure jobs that will disappear?
- No. It measures exposure, including work that AI may complement.
- Is AI already causing mass UK job losses?
- The evidence does not show that. In the ONS survey around half of businesses using AI report no change in headcount and few report cuts, and the government's review says the falls in hiring it describes do not yet establish AI as the cause.
- Does Microsoft's list predict job displacement?
- No. Its US preprint measured where chatbot use applies to occupations; it was not a displacement forecast.
Sources and checks
1.Up to 8 million UK jobs at risk from AI unless government acts · Institute for Public Policy Research
Published: 27 March 2024. Checked: 23 October 2026. Coverage: UK. Kind: scenario. Licence: © IPPR; short citation only.
- 7.9 million: jobs in IPPR's worst-case scenario (UK, scenario). An illustrative full-displacement scenario, not a prediction.
- 4.4 million: jobs in IPPR's central scenario (UK, scenario). An illustrative scenario with different adoption and policy assumptions.
- zero: job losses in IPPR's full-augmentation scenario (UK, scenario). An illustrative scenario, not a prediction.
- 1.5 million: jobs in IPPR's worst case for generative AI as it existed then (UK, scenario). The full-displacement case for the first wave of adoption; an illustration, not a prediction.
Supports: Illustrative UK displacement scenarios and task exposure. Conflict: None declared in this source record.
2.Assessment of AI capabilities and the impact on the UK labour market · Department for Science, Innovation and Technology and the AI Security Institute
Published: 28 January 2026. Checked: 23 October 2026. Coverage: UK. Kind: evidence review. Licence: Open Government Licence v3.0.
- around 70 per cent: UK workers in occupations with tasks AI could perform or enhance, per IMF estimates (UK, exposure index). IMF estimates as quoted in the government's review; includes work AI may enhance as well as work it may take over.
- around half: of exposed UK workers in high-complementarity roles, per the IMF (UK, exposure index). A judgement about the social and physical context of work; the review calls such indices indicative.
- around 20 months: after which UK job postings returned to their original levels, in the KCL working paper (UK, study finding). The government review's summary of the King's working paper; it adds that causation is not established.
Supports: The UK exposure and complementarity estimates it quotes from the IMF, its summary of the King's working paper, and its caution on causation. Conflict: None declared in this source record.
3.The Future of Employment · Oxford Martin School
Published: 1 September 2013. Checked: 23 October 2026. Coverage: US. Kind: working paper. Licence: © authors; short citation only.
- about 47 per cent: US employment estimated technically susceptible to computerisation (US, scenario). A historical technical estimate, not a count of lost jobs or a UK generative-AI forecast.
Supports: Historical US computerisation estimate. Conflict: None declared in this source record.
4.The probability of automation in England · Office for National Statistics
Published: 25 March 2019. Checked: 23 October 2026. Coverage: England. Kind: official statistic. Licence: Open Government Licence v3.0.
- 0.27: ONS automation score for programmers and software development professionals (England, official statistic). A score from zero to one in a task model built before today's generative-AI tools; not a forecast.
- 0.51: ONS automation score for electricians and electrical fitters, England, 2017 data (England, official statistic). A score from zero to one in a task model built before today's generative-AI tools; not a forecast.
Supports: England task-based automation scores, different from current generative-AI exposure. Conflict: None declared in this source record. Data file
5.Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence · Stanford Digital Economy Lab
Published: 12 August 2026. Checked: 23 October 2026. Coverage: US. Kind: working paper. Licence: © authors; short citation only.
- 19 per cent: relative employment gap for young US workers in exposed occupations (US, study finding). A revised relative gap in a payroll sample; the authors found no economy-wide displacement.
Supports: Revised US early-career payroll finding: first released in August 2025, cited here in its revision of 12 August 2026. Conflict: None declared in this source record.
6.New study reveals early impact of AI on job market in UK · King's College London
Published: 7 October 2025. Checked: 23 October 2026. Coverage: UK. Kind: press release. Licence: © King's College London; short citation only.
- 5.8 per cent: decline in junior positions at highly exposed firms in a UK study (UK, study finding). Reported in a university press release; the working paper itself could not be opened for this page.
Supports: UK university press release on junior roles. Conflict: None declared in this source record.
7.Artificial intelligence in UK businesses · Office for National Statistics
Published: 20 July 2026. Checked: 23 October 2026. Coverage: UK. Kind: survey. Licence: Open Government Licence v3.0.
- around half: of businesses using AI, of all sizes, reporting no impact on their headcount (UK, survey). Self-reported; many other businesses said the effect was uncertain or not applicable.
Supports: What UK businesses report about AI and headcount. Conflict: None declared in this source record.
8.Gen-AI: Artificial Intelligence and the Future of Work · International Monetary Fund
Published: 14 January 2024. Checked: 22 October 2026. Coverage: global. Kind: exposure index. Licence: © IMF; short citation only.
Supports: The original exposure and complementarity work. Conflict: None declared in this source record.
9.Working with AI: Measuring the Applicability of Generative AI to Occupations · Microsoft Research (arXiv preprint)
Published: 10 July 2025. Checked: 23 October 2026. Coverage: US. Kind: preprint. Licence: arXiv author submission; citation only.
Supports: US AI-applicability preprint, not displacement. Conflict: Microsoft develops and sells AI products.
10.Open Government Licence v3.0 · The National Archives
Published: undated. Checked: 22 October 2026. Coverage: UK. Kind: licence. Licence: Open Government Licence v3.0.
Supports: Attribution terms for public sector material. Conflict: None declared in this source record.
Contains public sector information licensed under the Open Government Licence v3.0. No publisher endorses TradeTraining or this guide.
Changelog
- : Added practical questions for interpreting the evidence and structured data for the visible FAQs.
- : Published after the research hosts were approved. Re-read every source; added IPPR's first-wave scenarios, the ONS automation model, ONS business survey results, the other half of the IMF complementarity split, IPPR's own warning and the government's summary of a later version of the King's paper; the UK exposure share is now quoted from the government's review because the IMF note could not be opened.
- : Drafted source-by-source fact check; publication awaits approval of research hosts.