Will AI replace my job?

Almost certainly not as a whole, and almost certainly in parts, and the parts matter far more than the whole. The best available estimates say roughly eighty percent of US workers could have at least ten percent of their work tasks affected by large language models, and about nineteen percent could see at least half their tasks affected. Meanwhile, the most rigorous measurement of what has actually happened to pay and hours in exposed occupations finds close to nothing, two years in. Both of those are true, and holding them together is the whole skill. Your job is being unbundled into tasks rather than deleted, and some of those tasks are being taken, and the question that decides your position is not how many go but which. If AI removes the least demanding parts of your role, what remains gets harder and you become more valuable. If it removes the most demanding parts, what remains gets easier, a lot more people can do it, and no amount of tool fluency will protect you.

Exposure, and what it actually measures#

On exposure, Eloundou, Manning, Mishkin and Rock produced the most widely used estimate, published in Science in 2024. Around eighty percent of the US workforce could have at least ten percent of their work tasks affected by large language models, and roughly nineteen percent of workers could see at least half of their tasks affected. About fifteen percent of all worker tasks could be done significantly faster at the same quality using a model directly, rising to somewhere between forty-seven and fifty-six percent once you account for software built on top of models. Read the wording carefully, because it is routinely misquoted: this is an estimate of task exposure, meaning the work could be materially assisted or accelerated. It is not a forecast of job losses, and the authors are explicit about that.

On what has actually happened, Humlum and Vestergaard linked large-scale adoption surveys to administrative labour records in Denmark, covering roughly 25,000 workers across 7,000 workplaces in eleven exposed occupations including accountants, journalists, legal professionals, software developers and marketers. Two years after ChatGPT launched they found precise null effects on earnings and recorded hours, ruling out effects larger than two percent. What did move was the structure of the work: task reorganisation, new tasks in content generation, AI oversight and AI integration, and adopters moving into higher-paying occupations. Their summary is worth memorising, because it answers the panic and the complacency at once: technological change reshapes work well before it surfaces in earnings or hours.

On which direction that reshaping runs, Autor and Thompson give the sharpest result available. Analysing four decades of task data across 303 US occupations, they found that automation which removed the less expert tasks from a job raised wages and reduced employment, while automation which removed the expert tasks lowered wages and increased employment. The same technology produces opposite outcomes for the humans depending on what it leaves behind, and expertise, not exposure, is the variable that determines which one you get.

On adoption, Bick, Blandin and Deming found that by late 2024 nearly forty percent of US adults aged 18 to 64 used generative AI and twenty-three percent of employed people had used it for work in the previous week, but that only one to five percent of all work hours were actually being assisted. And Brynjolfsson, Li and Raymond found the productivity gains concentrated among the newest and least experienced workers, at thirty percent, with almost no effect on the most skilled. AI raises the floor much faster than it raises the ceiling, which is good news for anyone starting and awkward news for anyone whose value rested on being better than a beginner.

The strongest case against this page#

David Autor makes the most rigorous optimistic argument in the field, and it deserves to be read rather than summarised away. In Applying AI to Rebuild Middle Class Jobs (2024), he argues that AI differs from previous waves of automation in a specific way: earlier technology automated routine work and hollowed out the middle, while AI can extend expert decision-making to workers who do not currently possess the expertise. On that reading AI could rebuild middle-skill, middle-income work rather than erode it further, by allowing more people to do work that currently requires years of training.

If Autor is right, the anxiety this page addresses is largely misplaced, and the correct response is to accelerate rather than protect. I take the argument seriously and I think it turns on one question the paper does not settle: whether extended expertise is acquired or merely borrowed. A nurse practitioner supported by AI to make decisions previously reserved for physicians is genuinely more capable if the support builds judgement over time, and is exposed if it substitutes for judgement she never develops. Autor's mechanism and the capability argument on this site are compatible only under the first condition, and nothing guarantees it.

Daron Acemoglu supplies the other corrective, from the opposite direction. In The Simple Macroeconomics of AI (2024, published in Economic Policy in 2025), he estimates AI's effect on total factor productivity over a decade as modest, roughly an order of magnitude below the most quoted forecasts. That is worth holding against every transformation headline, including some cited approvingly elsewhere on this site.

And the exposure picture looks different from outside the United States. The ILO's refined global index of occupational exposure (2025), updating its 2023 analysis, finds augmentation dominating automation in most contexts, with clerical work the most exposed category and a markedly larger effect on women in higher-income countries. The IMF's staff analysis adds the finding that advanced economies face more exposure and are better placed to benefit, which widens global gaps rather than narrowing them.

The shakiest number people quote#

The exposure estimates are the shakiest number people quote most confidently. They rest on human and model judgements about what tasks could be affected, using task descriptions from an occupational database, and they say nothing about whether a firm will adopt, whether the economics work, or whether the remaining work expands to fill the time. Treat them as a map of where the pressure falls, not as a countdown.

The Danish null result is the most credible measurement available and it comes from a high-trust, high-wage, heavily unionised labour market with strong employment protection, over a short window. It does not settle the question for a US technology firm or a UK professional-services partnership, and two years is early for a general-purpose technology. The Autor and Thompson data runs to 2018, so their model is a way of thinking about generative AI rather than a measurement of it.

Entry-level hiring is the one area where reasonable people read the same data in opposite directions, with some finding a real contraction in graduate roles in exposed occupations and others attributing it to interest rates and post-pandemic correction. That argument is set out separately in will AI replace entry-level jobs. Anyone telling you it is settled, in either direction, is ahead of the evidence.

A badly formed question#

The question "will AI replace my job" is badly formed, and badly formed questions produce bad answers. Jobs are not the unit that is moving. Tasks are. AI is unbundling roles: swallowing some tasks overnight, stretching and recombining others, and leaving a smaller human core under pressure to justify itself. I wrote about this in The Great Unbundling of Work in May 2025, and the sentence that travelled furthest was the plainest one: your job is not disappearing, it is dissolving.

Which means the thing to fear is the slower, quieter version I called hollowing, rather than the redundancy notice: expertise that took twenty years to build being commoditised out from under someone while the job title, the desk and the salary stay exactly where they are. Nobody announces it. There is no restructure to point at. The work simply becomes something a great many more people could do, and the market notices before the person does.

So replace the unanswerable question with a specific one, in three parts. Which of my tasks can a capable model do adequately today? Once those are gone, is what remains harder or easier than my job is now? And can I demonstrate that I can do the harder part, given that my output no longer proves it? That third question is the one people miss. It is becoming the important one. When anyone can produce competent-looking work, competent-looking work stops being evidence of anything, and the burden shifts to showing your judgement rather than your artefacts.

There is a trap on the way up, too. AI lets someone early in a career produce output that looks like it carries fifteen years of judgement when it carries eighteen months, and the gap only appears when a decision arrives that the model cannot make. I call that synthetic seniority. It is a fast route to promotion and a slow route to being stranded, because you arrive in the senior role having skipped the repetitions that were supposed to prepare you for it.

Assessing your own exposure#

Twenty minutes, honestly done, is worth more than any list of future-proof jobs.

• Write down last week as tasks, not as a role. Everything you actually did, in units of an hour or less. Most people have never seen their job written this way.
• Mark each one. A model can do this now; a model will do this within two years; this needs a human, and say why.
• Delete the first category and read what is left. Is this a harder job than the one you have, or an easier one? That answer is your position, and more informative than any industry forecast.
• Find your frontier. Name the specific situations in your domain where the model sounds most confident and is most often wrong. That knowledge does not transfer and does not commoditise, and the jagged-frontier research says almost nobody has it.
• Check whether your judgement is visible. If someone had to distinguish your work from a competent person using the same tools, what would they look at? If the answer is nothing, that is the gap to close.

Move towards the harder side#

Move towards the harder side of your role deliberately. The ambiguous, judgement-heavy, politically awkward work that nobody volunteers for is the work that holds its value. It is usually available.

Do not build a career on tool fluency. Interfaces are being engineered to need less skill every quarter. Knowing the tools is table stakes; knowing when they are wrong is not.

Keep taking the reps you could now skip. The capability that carries your value should be exercised unaided at intervals, for the same reason pilots fly manual approaches. See using AI without dependency.

Make your reasoning visible. Decision logs, recorded rationale, the annotated draft showing what you prompted and what you changed and why. This is how you avoid being priced as though the machine did it.

Stop waiting for a signal. The pay data is the slowest indicator there is. By the time it moves, the positions will have been taken. The fuller version of the response is in staying valuable in the age of AI.

Development of the idea#

The Great Unbundling of Work (25 May 2025) set out the tasks-not-jobs argument, the hollowing of expertise, and the five patterns of resistance people show when their role is being rebuilt. Knowledge Is No Longer Power developed the argument about where the human edge moves once information is free. In the Observer in July 2026 I argued that the next AI power class will not build the models. The longer treatment is in SuperSkills (Kogan Page, 2026).

Key research and primary sources

• Gmyrek, P., Berg, J. and Bescond, D. (International Labour Organization) (2023). Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality. ILO Working Paper 96, 21 August 2023.
• Massenkoff, M. and Huang, S. (Anthropic) (2026). What 81,000 people told us about the economics of AI. Anthropic, 22 April 2026.
• Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2024). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. Published in Science, 384(6702), 1306-1308.
• Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777, revised March 2026.
• Autor, D. and Thompson, N. (2025). Expertise. NBER Working Paper 33941; published in the Journal of the European Economic Association, 23(4).
• Bick, A., Blandin, A. and Deming, D. J. (2024). The Rapid Adoption of Generative AI. NBER Working Paper 32966.
• Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889-942. Earlier version NBER Working Paper 31161.
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School and BCG working paper.
• World Economic Forum (2025). The Future of Jobs Report 2025.

Related SuperSkills research#

On what to do next, staying valuable in the age of AI and human skills in the age of AI. On early careers, will AI replace entry-level jobs and the missing rungs. On the traps, synthetic seniority and the verifier's discount. On what remains distinctly human, what stays human. On the advice everyone gives, see why "learn to prompt" is weak career advice. The graded evidence is in the evidence base. See which jobs are safest from AI.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This work draws on research across more than 200 organisations in 30 countries over seven years. Findings are attributed to the studies that produced them and kept separate from the interpretation, which is the author's. Task exposure, unbundling and the task-versus-job distinction are established ideas in labour economics and are not his coinages; synthetic seniority, the verifier's discount, the missing rungs and capability debt are part of the SuperSkills lexicon. This is a living reference on a 90-day review cycle, given how quickly the labour-market evidence is moving.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.