Most of the alarming numbers about AI and jobs count tasks that software could reach, not jobs that are about to vanish. So when you ask “will AI take my job?”, the best research so far suggests that for most people AI is more likely to change what fills the working week than to end the job, though direct evidence so far is thin. The 2025 global index from the International Labour Organization (ILO) puts one in four workers worldwide in occupations with at least some exposure to generative AIgenerative AI: AI systems that produce new text, images, audio or code in response to a request, by generating output that resembles the data they were trained on. Chatbots built on large language models are the best-known kind.Full entry in the glossary, and only about 3 percent in the most exposed group.¹ Since nearly every occupation mixes tasks a model can handle with tasks that need a person, the index’s authors judge that reshaping jobs, rather than removing them, is the likeliest result.¹

Two habits make this research usable: read every AI-jobs figure for what it counted, and judge your exposure by the tasks in your week, not your job title. If the headlines already have you weighing a career change, that decision has its own guide.

What an AI exposure score actually measures

An AI exposure score estimates how much of a job’s task list an AI system could technically do or speed up, as judged by human raters or by an AI model. It does not measure whether employers will adopt the technology, whether the job will disappear, or when anything will happen.

Researchers rate each task on an official list, such as the US O*NET database or the ILO’s international ISCO classification, average the ratings for each occupation, then weight them by how many people hold each job. No step in that chain observes what employers actually do, so the result describes what is technically possible, not what anyone has decided.

1. 1Occupationfrom an official list, such as O*NET or ISCO
2. 2Task listwhat people in that job are described as doing
3. 3Rate each taskcould AI do it or speed it up?
4. 4Share of workersscores weighted by how many people hold each job

The best-known version came from a team at OpenAI and the University of Pennsylvania.

The study

Limited evidence

Every US task rated for language-model reach, 2023 working paper

The team counted a task as exposed if access to a large language model, or to software built on one, could cut the time needed to do it by at least half at the same quality. On their main measure, which half-weights tasks that need extra software, about 80% of US workers were in occupations with at least a tenth of tasks exposed, and about 19% in occupations with at least half. The authors say their measure does not distinguish between AI that helps workers and AI that replaces them, and they make no predictions about when the technology will be adopted.²

The authors flag their own main weakness: their raters knew language models well but came from few occupations, so they could misjudge what a task involves. Exposure was generally higher in better-paid occupations. The 2024 Science version frames its headline differently: by its abstract, about 2 percent of jobs could have over half their tasks affected by language models alone, and just over 46 percent once current and likely future software is counted.²

As an illustration, an accounts assistant might draft payment reminders, match invoices to orders, calm an angry supplier and chase a manager for sign-off. The first two could be rated exposed; the others run on relationships and authority. A partly exposed score shows where change may arrive first, not that the role will end.

Notwithstanding these findings, we stress that such exposure does not imply the immediate automation of an entire occupation, but rather the potential for a large share of its current tasks to be performed using this technology.

So translate “a share of jobs is exposed” as “some tasks inside those jobs could be done faster by software”, then ask what happens to the rest of the work.

Help or replacement: the same score can point either way

A high exposure score can mean AI helps you do your job or that it takes tasks over so fewer people are needed. The ILO and IMF analyses both say which one happens depends on the type of work and on employers’ choices.

In the ILO’s 2023 estimates for high-income countries, more than twice as many jobs had potential for augmentation as for automation.³ The International Monetary Fund (IMF) draws a similar line. Its examples of highly exposed jobs where AI is likely to support workers who stay responsible for decisions are surgeons, lawyers and judges; telemarketers are its example where AI is more likely to replace tasks and demand for workers could fall.⁴

The IMF notes that even where AI complements the work, people can be displaced if they lack the skills to use it or their employer does not invest.⁴ The ILO adds a warning from the other direction: a job that keeps its human tasks may still need fewer people if the time saved lets a smaller team do the same work.¹

As an illustration: a tool can draft a contract clause, but a lawyer signs off and answers for it; a tool can run a sales script, and nobody needs to stand behind each call.

Sort your own week before you trust a headline

The most useful exposure estimate for you is one you build from your own task list rather than your job title, because, as a 2016 OECD analysis of 21 countries stresses, workers in the same occupation often perform different tasks.⁵

1. List what you did last week, 10 to 20 tasks, each with a verb (“drafted the monthly client update”).
2. Sort each task into three groups: AI could draft or do most of it; AI could help, but you check, decide or answer for it; or it needs your presence, relationships or authority.
3. Estimate the share of your time in each group. Most hours in the first group means more exposure than your title suggests.
4. Ask whether your employer is already trying AI tools on your first-group tasks.

As an illustration, take a customer-support team lead. Drafting reply templates and summarizing the week’s tickets go in the first group; approving refunds a tool has suggested goes in the second, because the lead answers for the money; coaching a new hire and calming an escalated customer go in the third. If the first group fills a few hours a week, the job is far less exposed than a headline about “customer service” implies. If it fills half the week, that half is the part of the role to watch and to raise with your manager.

WiserHours

Seven days of task notes

• List 10 to 20 tasks from last week, each with a verb
• Sort them: AI could do it, AI could help while I check, or it needs me
• Estimate the share of my time in each group
• Find out which tasks my employer is already trying AI on

Find this checklist online:
https://wiserhours.com/career-change/will-ai-take-my-job/#checklist-will-ai-take-my-job-audit

If you manage a team, the same sorting underpins adding AI to shared team workflows, and your third-group tasks are a good start for spotting the transferable skills you already have.

Will AI take my job? Why the big estimates disagree

The best-known estimates of AI’s reach range from about a quarter of workers worldwide to a majority of jobs in rich economies, because they measure different things: which AI, who rated the tasks, how high the bar for “exposed” sits, and whether anything was measured at all.

Start with which AI. The IMF’s 2024 staff discussion note used a broad index that matches 10 AI applications against the abilities each job needs, covering AI in general rather than only generative AI, and stresses that its measures only rank occupations against each other.⁴ A wider definition can sweep in more tasks, so the same economy may look more exposed.

Then look at who did the rating. The ILO’s first global study, in 2023, had GPT-4 score the tasks in the ISCO classification and called its result an upper-bound estimate; even so, only clerical work came out highly exposed.³ The 2025 update, built with NASK, Poland’s national research institute, added a survey of workers in Poland and international expert review, and its global estimates came out lower.¹

Finally, some famous figures are not measurements. The World Economic Forum’s Future of Jobs Report 2025 is a survey of over 1,000 leading employers in 55 economies about what they expect by 2030, so its figures are expectations, not outcomes.⁶

The height of the bar can flip a headline on its own. Picture a job with ten tasks, three of which AI could do in half the time. A study that counts a job once a tenth of its tasks are exposed puts this worker in the exposed group; one that asks for half or more leaves them out. Same job, opposite headlines.

The famous Oxford estimate, and why researchers now count tasks

The most quoted figure in this debate, that 47 percent of US employment was at high risk of computerisation, comes from a 2013 Oxford working paper by Carl Benedikt Frey and Michael Osborne. Its authors said they made no attempt to estimate how many jobs would actually be automated.⁷

Working with machine-learning researchers, they labelled 70 occupations by hand as automatable or not, largely from their task descriptions, and let an algorithm extend those judgements to 702 US occupations; at risk meant automatable “perhaps over the next decade or two”.⁷ Label a whole occupation automatable and everyone in it counts, including people who spend much of their day on tasks that are hard to automate.

Myth

Oxford researchers predicted that almost half of all jobs would vanish.

Fact

Frey and Osborne estimated the share of US employment in occupations at high risk of computerisation, and said they made no attempt to estimate how many jobs would actually be automated.

The main critique came in a 2016 OECD working paper by Melanie Arntz, Terry Gregory and Ulrich Zierahn: labelling whole occupations overstates the risk, because high-risk occupations often still contain many tasks that are hard to automate. Once they allowed for differences in tasks between workers in the same occupation, about 9 percent of jobs looked automatable on average across 21 OECD countries. They also noted that experts tend to overestimate new technology, especially for tasks needing flexibility, judgement and common sense.⁵

1. Counting whole jobs: the job is labelled at risk or safe as one unit
2. Counting tasks: each task is rated on its own: teal, AI could do much of it; amber, AI could help while a person checks or decides; plain, it needs people (tile counts are illustrative)

Think of two bookkeepers. One keys receipts into software all day; the other spends half the week explaining the numbers to a small-business owner. Same occupation code, but a receipt-reading tool changes the first job far more. So treat the 47 percent figure as a 2013 statement about occupation labels, and judge your own position by what you do between Monday and Friday.

What has actually changed for workers so far

Direct evidence on what AI has done to jobs and pay is early and mixed. A large Danish study found no measurable change in earnings or hours about two years after ChatGPT’s launch, while a US payroll study found employment of the youngest workers in AI-exposed occupations falling behind their peers. Neither working paper has been peer reviewedpeer review: The checking of a study by independent experts, usually arranged by a journal, before it is accepted for publication. It screens for weak methods and unclear reporting, but reviewers rarely see the raw data, so passing it does not prove a finding is right.Full entry in the glossary yet.

In Denmark, Anders Humlum and Emilie Vestergaard linked adoption surveys to official employment records. Their abstract reports that most employers in exposed occupations had launched chatbot initiatives, yet rules out effects on earnings or recorded hours larger than 2 percent; what shifted was how work was organized, with new tasks such as overseeing AI output.⁸

In the US, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen of Stanford track records from the payroll company ADP in a working paper revised in August 2026. Employment of workers aged 22 to 25 in AI-exposed occupations stands about 19 percent below where it would be had it kept pace with less-exposed peers, mostly through less hiring; experienced workers show no comparable gap. The authors call these early descriptive signals, not causal estimates: the gap shrinks when education is taken into account, and some of the divergence predates generative AI.⁹

The two results are less contradictory than they look: one tracks pay and hours in Denmark, the other head counts by age in the US. One possible mechanism the Stanford team offers is that AI substitutes more easily for codified, textbook knowledge, the core of formal education, than for the tacit know-how that comes with experience.⁹ Picture, as an illustration, a first-year analyst whose early months went on summarizing reports and building standard slides: if a tool drafts most of that, the starter tasks shrink before any job title changes. If you are early in your career, or hire people who are, watch how entry-level roles in your field are changing.

The bottom line

A figure about AI exposure counts tasks software could reach, not jobs that will go, and the ILO expects most exposed jobs to change rather than disappear. Evidence on actual job losses is early and mixed, with the clearest warning so far, from one US study, for entry-level workers in exposed fields. Your best estimate comes from a week of your own task notes, not from a headline.