Will AI replace jobs, and which ones?

The question arises occupation by occupation: accountants, developers, doctors, lawyers, teachers, journalists. The studies that answer it measure an exposure : the share of an occupation's tasks that a generative AI model could perform or assist. The International Labour Organization (ILO), in its May 2025 update, scores each task from 0 to 1, then averages by occupation. This score does not say whether the tasks will be given to AI, or whether the occupation will lose jobs; the authors of the ILO working paper consider the transformation of occupations more likely than their disappearance, as most keep tasks that require a person; they add that even partial automation can reduce labour demand in the most exposed occupations.

This page places side by side, for 12 occupations, this score and the change in US employment from 2022 to 2025, a country whose public statistics give employment by occupation at the end of each year. It makes no forecast.

In the 6 occupations that the ILO places in gradient 3, US employment moved in 2 directions: it fell by 22% among secretaries and by 4% among customer service representatives, and rose by 8% among developers and by 7% among accountants, faster than total employment (+3%). In the not exposed occupations, it also varied: +5% among nurses, +2% among doctors. The 21% fall among secondary school teachers is mainly due to 2025, a year in which the neighbouring line for other teachers and instructors goes from 805,000 to 1,018,000: it may come from the classification of responses more than from employment. Over 3 years, employment in an occupation depends on the economic cycle, interest rates, public budgets and the survey itself as much as on AI; this table does not allow a change to be attributed to it.

The ILO finds the highest exposure in office jobs: data entry, typing, routine accounting, secretarial work. Since its 2023 version, it has raised the exposure of several skilled occupations, including financial analysts, web and multimedia developers, translators and applications programmers. Occupations that require physical presence, care or direct contact with people, such as nurses, come out as not exposed.

Actual use follows this map in part. In the conversations analysed by Anthropic, the publisher of Claude, in November 2025, computing and mathematics account for about a third of uses, education 15%. On the consumer app, 52% of exchanges are used to work with AI and 45% to hand it the whole task; in programmatic uses, businesses hand it the whole task in about 3 cases out of 4. Our analysis of productivity at work says how much time these uses save.

The only decline that studies link to exposure concerns young people entering occupations. In France, permanent-contract (CDI) hires of 21-29 year olds fell by 28% from late 2022 to March 2026 in the most exposed occupations, where the study places translators and journalists, against 8% in the least exposed, where it places nurses. For developers and computer engineers aged 21-29, the decline reaches 42%. In the United States, payroll data studied at Stanford show a gap in the same direction in the salaried employment of 22-25 year olds. None of these studies establishes the cause of the decline. Our analysis of early-career workers details these studies, their limits and other countries.