Will AI replace my job? What the exposure research says for your role

The honest answer to "will AI replace my job" has three parts, and most pages give only one. Exposure studies measure which tasks a model could do; they are not forecasts of job losses, and their authors say so. The field evidence from 2023 to 2026 measures what has actually happened, which is less than either camp predicted and unevenly distributed. And any statement about 2030 is a scenario with a band, including ours. The query is small, about 210 US searches a month, but growing 91 % year on year and 24 % a year over three years (DataForSEO, retrieved 2026-11-09).

Exposure vs replacement

Every headline number in this field comes from one of two methods. The first scores occupations by how many of their tasks a technology could perform. The second counts what happened to employment and wages. They are routinely confused, and the confusion runs in one direction: exposure gets reported as displacement.

The best-known exposure estimate is Frey and Osborne's, first circulated in 2013: 47 % of US employment in occupations at high risk of computerisation over "a decade or two" B^[1]. A decade later, US unemployment was near historic lows. The estimate was not wrong about tasks; it treated whole occupations as the unit. When OECD researchers rescored the same question at the task level, since most jobs mix automatable and non-automatable tasks, the share of jobs at high risk fell to about 9 % B^[2].

The generative-AI estimates learned from this. Eloundou and colleagues, scoring tasks against large language model capabilities, find that around 80 % of the US workforce could have at least 10 % of their tasks affected and about 19 % could have half or more; exposure rises with wages and is highest in information processing and lowest in manufacturing, agriculture and mining A^[3]. The authors are explicit that "affected" means the task could be done at least 50 % faster with the model's help, not that it disappears.

The ILO's refined global index is the most careful of the official measures. It places about one in four workers worldwide in an occupation with some generative-AI exposure, 34 % of employment in high-income countries against 11 % in low-income ones A^[4][5]. Only 3.3 % of global employment falls into the highest exposure tier, where many tasks have high automation potential, and women's employment is roughly twice as likely as men's to be in it (4.7 % against 2.4 %), because clerical work is both highly exposed and female-dominated A^[6]. The ILO's own framing is augmentation for most and transformation for few. The OECD reaches a similar shape: about a quarter of workers in exposed occupations, roughly 1 % highly exposed, though we hold that record at grade C because it was read through a secondary summary C^[7].

The headline numbers and their ranges

Figure	Source	What it measures	Grade
47 % of US employment at high risk	Frey & Osborne 2013/2017 [1]	Occupation-level automatability, pre-LLM	B
9 % of jobs at high risk across OECD	Arntz, Gregory & Zierahn 2016 [2]	Same question at task level	B
80 % of US workers with 10 %+ of tasks affected; 19 % with 50 %+	Eloundou et al. 2024 [3]	LLM task exposure, not job loss	A
25 % of global employment exposed; 3.3 % in the highest tier	ILO 2025 [4][6]	Occupational exposure gradient	A
170 million jobs created, 92 million displaced by 2030 (net +78 million, 22 % churn)	WEF Future of Jobs 2025 [8]	Employer expectations, 1,000+ firms, extrapolated	A
39 % of workers' skills transformed or obsolete by 2030	WEF 2025 [9]	Employer expectations	B
59 of 100 workers need reskilling by 2030, 11 unlikely to get it	WEF 2025 [10]	Employer expectations	A
Total-factor-productivity gain of at most about 0.66 % over ten years	Acemoglu 2024 [11]	Macro model built on the exposure estimates	B

Read the table as a set of brackets: a large share of tasks could change; employers expect more jobs created than displaced, with heavy churn; the aggregate productivity effect at current capabilities is modest. None of these is a count of jobs that have ended.

What actually happened, 2023 to 2026: the early evidence

The controlled studies measure productivity, and they are consistent about one thing: effects depend on the task, the worker's experience and whether the task sits inside the model's competence. Across 5,172 customer-support agents, a generative-AI assistant raised issues resolved per hour by about 14 to 15 % on average, over 30 % for the least experienced and close to zero for the most experienced A^[12]. In a preregistered experiment with 453 professionals, ChatGPT cut time on writing tasks by about 40 % and compressed the quality gap between weaker and stronger writers A^[13]. Among 758 consultants, AI improved quality on tasks inside its competence and made people 19 percentage points less likely to be correct on a task just outside it, without their noticing B^[14]. Sixteen experienced open-source developers took 19 % longer with AI tools on their own repositories while believing they were faster B^[15]. And across 106 experiments, human–AI combinations on average under-performed the better of human or AI alone on decision tasks, while helping in content creation A^[16]. There is no single percentage.

The labour-market evidence is newer and thinner, which is why it carries B grades. In Denmark, matching surveys of 25,000 workers in 7,000 workplaces to administrative records across eleven exposed occupations, Humlum and Vestergaard find rapid adoption, average time savings of a few percent of work hours, and no detectable effect on earnings or recorded hours two years in; their confidence intervals rule out effects larger than about 1 % B^[17]. In the United States, using payroll records from millions of workers, Brynjolfsson, Chandar and Chen find that employment for early-career workers aged 22 to 25 in the most AI-exposed occupations fell by about 13 % relative to less-exposed occupations since late 2022, while employment for more experienced workers in the same occupations, and overall employment, continued to grow B^[18]. Both are working papers and both describe the first three years of a technology, not its equilibrium.

Adoption is broad but shallow. A third of US adults have ever used ChatGPT, roughly double the 2023 share A^[19]; almost half of employed people in Germany use AI at work, but only 8 % daily B^[20]. The pattern the data describe so far is augmentation at the task level, early-career entry pressure in the most exposed occupations, and no aggregate displacement.

A self-assessment by task type

The useful unit is not your job title but the tasks your week is made of. Score each row for the share of your week it represents, then read the two columns on the right.

Task type	Exposure signal in the research	What the field evidence shows so far	Reading
Routine drafting, summarising, translation	Highest LLM exposure in the task scoring [3]	Large time savings, quality floor raised [13]	Tasks shrink; the job that is only this is at risk
Structured data entry and clerical processing	ILO highest tier is dominated by clerical work [6]	Where it is automated, it goes; early-career clerical entry roles under pressure [18]	Highest risk by volume
Scripted customer support	Direct field study [12]	Output per agent up 14–15 %, most for novices; fewer novices needed per unit of demand	Fewer entry roles, more supervision roles
Analysis with judgement under uncertainty	Exposed in task scoring, but human–AI combinations lose on decision tasks [16]	Gains inside the model's competence, errors outside it [14]	Durable if you own the check, not the draft
Software development	High exposure [3]	Novices faster on bounded tasks [12]; experienced developers slower on their own code [15]	Task mix changes; verification becomes the job
Physical, hands-on and in-person work	Lowest exposure in every index [3][4]	No measurable LLM effect	Low risk from this technology
Relationship, negotiation and care work	Low to moderate exposure [3]	Not measured in the controlled studies	Low risk; augmentation at the edges
Verification, acceptance and oversight of AI output	Not an occupation in the indices; a task that grows with adoption	Automation bias is real and falls with training and accountability [21]	The growth area in every scenario

If most of your week sits in the first three rows, the realistic risk is not that your job vanishes next year but that the number of people needed for it falls and that the entry roles disappear first, which is what the payroll data show ^[18]. If most of it sits in the bottom four, the risk is of using the tools badly rather than of being replaced by them.

Our radar scenarios

Our Future Demand Radar scores twelve productivity categories from keyword histories, Google Trends shares, graded evidence and explicit assumptions, then runs 4,000 Monte Carlo draws per category. The category this page sits in, future-of-work skills and AI literacy, scores 45.8 out of 100 in the baseline scenario (p10–p90 band 41.4–50.2), which places it in the "monitor" class; the conservative scenario gives 42.5 ("monitor") and the accelerated scenario 48.2 ("validate in 12 months"). The numbers below are multipliers of today's US search demand for the category, not jobs.

Scenario	What it assumes about jobs	Skills demand 2028 / 2030 / 2035
A, conservative	Copilots stay features inside existing suites; task change without role change; exposure tiers unchanged	1.2× / 1.4× / 1.5×
B, baseline	Delegated agents routine in knowledge work by 2028–2030; entry roles in the most exposed occupations thin out; new roles form around specifying, verifying and orchestrating; AI Act literacy duties sustain training demand	1.5× / 2.2× (band 1.7–2.8) / 2.8× (band 2.0–4.0)
C, accelerated	Capable agents absorb much routine knowledge work; the ILO highest tier rises with observed displacement; verification and judgement become the core of most roles	1.7× / 3.0× / 4.4×

Our baseline scenario expects task-level change for most knowledge workers, visible displacement concentrated in the ILO's highest tier and in entry-level roles, and net job counts that continue to depend on the economy more than on the technology, consistent with the Danish and US evidence to date ^[17][18]. We do not plan on scenario C, because the organisational-impact data have not moved in its direction yet, and we check the signposts quarterly: the ILO tier share, the early-career employment gap in exposed occupations, and the share of firms reporting measurable returns. Forecasts are model outputs with stated uncertainty, not facts.

How to lower your risk

The evidence supports a short list, and it is not "learn to prompt".

1. Move your week toward the bottom half of the table. Take on the checking, the specification and the client-facing parts of your current work; let the drafting go to the tool and own the acceptance.
2. Become the person who catches the wrong answer. Automation bias, accepting a confident wrong output, is measurable and falls with training and accountability A^[21]. In the consulting experiment the costly errors came from trusting the tool outside its competence ^[14]. Someone has to know where the edge is; make it you.
3. Pair the tool with a domain. The durable premium in the studies accrues to judgement about the domain, not to tool fluency, because the tool raises the floor for everyone ^[12][13]. Our ranking of AI skills for 2030 rates skills by whether they survive a change of vendor.
4. If you are early in your career, choose the exposed occupation's less exposed tasks. The payroll evidence shows the pressure landing on entry roles in the most exposed occupations ^[18]. Seek roles where you are supervising and checking output, not producing the first draft.
5. Treat the human skills as skills. Coordination, negotiation and care have the lowest exposure in every index ^[3][4]; our page on human skills in an AI economy covers what the evidence says about each, and jobs of the future covers where the employer surveys expect growth.
6. Build the system, not the habit. Deciding which tasks you delegate and which you verify is the operating decision that most changes your exposure.

What to do next

Take your task-type scores from the table to the OS Builder. It asks what your week is made of, which tasks you already delegate and how you check them, and returns a configuration that moves routine production to the tools and keeps specification and verification with you, using the same evidence grades as this page. We re-pull the exposure indices, the payroll and Danish follow-ups and the WEF and ILO releases every quarter; subscribe below for the quarterly summary.

Frequently asked questions

Will AI take my job?

For most people the evidence so far points to tasks changing rather than jobs vanishing. The ILO places about one in four workers in exposed occupations but only 3.3 % in the highest exposure tier ^[4][6], and the early labour-market studies find no detectable aggregate effect on earnings or hours ^[17], with pressure concentrated on entry-level roles in the most exposed occupations ^[18]. The risk is highest if your week is mostly routine drafting, clerical processing or scripted support.

What does AI exposure mean, and how is it different from job loss?

Exposure scores how many of an occupation's tasks a technology could perform or speed up. Eloundou and colleagues count a task as exposed if a language model could do it at least 50 % faster ^[3]. It is not a forecast of displacement: Frey and Osborne's 47 % exposure estimate from 2013 was followed by a decade of low unemployment, and rescoring at task level gave 9 % ^[1][2].

Which jobs are most at risk from AI?

By volume, clerical and administrative work, which dominates the ILO's highest exposure tier and is why women's employment is about twice as exposed as men's ^[6]. By task, routine drafting, summarising, structured data processing and scripted customer support show the largest measured productivity effects ^[12][13], which means fewer people are needed per unit of demand. Physical, in-person and relationship work scores lowest in every index ^[3][4].

Has AI already caused job losses?

Not in aggregate, as far as the early studies can tell. In Denmark, two years of adoption produced no detectable effect on earnings or hours across eleven exposed occupations ^[17]. In the United States, payroll data show a relative employment decline of about 13 % for workers aged 22 to 25 in the most exposed occupations since late 2022, while older workers in the same occupations and total employment kept growing ^[18]. Both are working papers covering the first three years.

What is your forecast?

Our baseline scenario expects task-level change for most knowledge workers, displacement concentrated in the highest exposure tier and in entry-level roles, and net job counts driven more by the economy than by the technology. In our radar the future-skills category scores 45.8 out of 100 in that scenario (band 41.4–50.2), with search demand for skills reaching about 2.2 times today's level by 2030 (band 1.7–2.8). It is a model output with stated uncertainty, revisited whenever we re-run the model.

What should I learn to protect my job from AI?

Not prompting. The durable skills in the studies are task selection (knowing where the model's competence ends), verification of output, specification of delegated work and domain expertise, because gains accrue inside the model's competence and errors outside it ^[12][14][16], and because automation bias falls with training and accountability ^[21]. Our AI skills for 2030 page ranks them by durability.

Sources and evidence

1. The future of employment: How susceptible are jobs to computerisation? — Frey, C. B.; Osborne, M. A. (2017), observational B doi.org/10.1016/j.techfore.2016.08.019 (retrieved 2026-11-09)
2. The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis (OECD Social, Employment and Migration Working Papers No. 189) — Arntz, M.; Gregory, T.; Zierahn, U. (2016), report A doi.org/10.1787/5jlz9h56dvq7-en (retrieved 2026-11-09)
3. GPTs are GPTs: Labor market impact potential of LLMs — Eloundou, T.; Manning, S.; Mishkin, P.; Rock, D. (2024), observational A doi.org/10.1126/science.adj0998 (retrieved 2026-11-09)
4. Generative AI and Jobs: A Refined Global Index of Occupational Exposure (ILO Working Paper 140), global exposure share — International Labour Organization (2025), report A www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-o (retrieved 2026-11-09)
5. Generative AI and Jobs: A Refined Global Index of Occupational Exposure (ILO Working Paper 140), exposure by country income group — International Labour Organization (2025), report A www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-o (retrieved 2026-11-09)
6. Generative AI and Jobs: A Refined Global Index of Occupational Exposure (ILO Working Paper 140), highest exposure tier and gender gap — International Labour Organization (2025), report A www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-o (retrieved 2026-11-09)
7. Who will be the workers most affected by AI? (OECD Artificial Intelligence Papers No. 26) — OECD (2024), report C www.oecd.org/content/dam/oecd/en/publications/reports/2024/10/who-will (retrieved 2026-11-09)
8. Future of Jobs Report 2025 – press release (French edition), job creation and displacement — World Economic Forum (2025), report B reports.weforum.org/docs/WEF_Future_of_Jobs_2025_Press_Release_FR.pdf (retrieved 2026-11-09)
9. Jobs market at a crossroads: which are the fastest growing and declining jobs? (summarising WEF Future of Jobs 2025) — Euronews (2025), report B www.euronews.com/business/2025/02/01/jobs-market-at-a-crossroads-which (retrieved 2026-11-09)
10. Future of Jobs Report 2025 – press release (French edition), reskilling need — World Economic Forum (2025), report B reports.weforum.org/docs/WEF_Future_of_Jobs_2025_Press_Release_FR.pdf (retrieved 2026-11-09)
11. The Simple Macroeconomics of AI (NBER Working Paper 32487) — Acemoglu, D. (2024), report B doi.org/10.3386/w32487 (retrieved 2026-11-09)
12. Generative AI at Work — Brynjolfsson, E.; Li, D.; Raymond, L. (2025), observational A doi.org/10.1093/qje/qjae044 (retrieved 2026-11-09)
13. Experimental evidence on the productivity effects of generative artificial intelligence — Noy, S.; Zhang, W. (2023), RCT A doi.org/10.1126/science.adh2586 (retrieved 2026-11-09)
14. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua, F.; McFowland III, E.; Mollick, E. R.; et al. (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
15. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Becker, J.; Rush, N.; Barnes, E.; Rein, D. (METR) (2025), RCT B doi.org/10.48550/arXiv.2507.09089 (retrieved 2026-11-09)
16. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro, M.; Almaatouq, A.; Malone, T. (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
17. Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI (NBER Working Paper 33777) — Humlum, A.; Vestergaard, E. (2025), report B doi.org/10.3386/w33777 (retrieved 2026-11-09)
18. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence — Brynjolfsson, E.; Chandar, B.; Chen, R. (Stanford Digital Economy Lab) (2025), report B digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six- (retrieved 2026-11-09)
19. 34% of U.S. adults have used ChatGPT, about double the share in 2023 — Pew Research Center (2025), report A www.pewresearch.org/short-reads/2025/06/25/34-of-us-adults-have-used-c (retrieved 2026-11-09)
20. Ein Drittel nutzt KI mindestens einmal pro Woche (Presseinformation) — Bitkom e.V. (2026), report B www.bitkom.org/Presse/Presseinformation/Ein-Drittel-nutzt-KI-mindesten (retrieved 2026-11-09)
21. Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman, R.; Manzey, D. H. (2010), review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)