What happens to entry-level workers when AI does their work?

In November 2022, the month ChatGPT was opened to the public, private companies in metropolitan France, excluding agriculture, temporary work and household employers, were signing a trend level of about 22,000 permanent contracts a month with young people aged 21 to 29 in the occupations most exposed to AI: translators, journalists, economic research analysts, bank operations technicians, developers. By March 2026, this trend was about 6,200 lower.

These figures come froma study by HEC Paris and the Banque de France↗, published in October 2026 using companies' mandatory payroll filings, and from its presentation by HEC↗. Our analysis of productivity regarded it, together with the Stanford study, as the only measured signal of AI's effect on employment. Their authors, like those of the other employment studies cited here, describe a gap without establishing its cause.

The study sorts 310 occupations into 5 groups of equal size, based on the score a language model gave their skills, taken from the job profiles of France Travail, the French public employment service: what generative AI could do or help with, given the physical presence and tacit knowledge they require. This score measures a capability, not how companies actually use AI. The study tracks every hire and every termination of a permanent contract from January 2019 to March 2026.

Among 21-29 year-olds, permanent-contract hires fell by 28.3% in the most exposed group and by 7.5% in the least exposed, which brings together care and physical-presence occupations: nurses, care assistants, childminders. The gap remains large up to age 40 and almost disappears after age 50.

Exits fell as well: contract terminations among 21-29 year-olds fell by 23% in the most exposed occupations and by 6% in the least exposed. The authors describe a "freeze in movements" (our translation), with fewer entries, fewer exits and teams that turn over less.

The clearest case is software developers and engineers. In early 2026, their permanent-contract hires of 21-29 year-olds were about 42% below their November 2022 level, compared with 11% in the 4 less exposed occupation groups. The decline also affects older developers, and the authors cannot separate it from the IT-specific correction after the pandemic. Excluding developers, a gap of 17 points remains between the most and least exposed occupations.

About 6,000 companies answered a Banque de France survey on their use of AI in early 2026, and their answers could be matched with their hiring. The nearly 600 that report intensive use have seen their permanent-contract hiring of young people develop less favourably since late 2022 than at other companies in their sector; in 2025, the gap is 17% to 25% depending on the quarter, according to the HEC presentation. The authors do not know whether these companies were already using AI when the gap appeared.

A Stanford team tracks the payroll records of millions of US employees. Its overall figure is set out, with its caveats, in our article on AI and employment↗ ; its dashboard↗ lets you look at the data occupation by occupation.

In exposed occupations as in the others, the number of experienced employees is rising; only the employment of the youngest is falling. The Stanford authors put forward an explanation they do not demonstrate: AI may mainly be taking over the verifiable, procedural tasks, based on classroom and textbook knowledge, that justified hiring entry-level workers. This link fades when they control for education, our article notes.

2 Harvard economists identified, among more than 280,000 US companies, those that were recruiting a specialist to integrate generative AI. 6 quarters after the start of 2023, junior employment in those companies was about 9% lower than in the others, with no break for senior staff. In their job postings, tasks that AI can do appear less and less often in junior positions after 2023, with no comparable change in senior positions.

Other teams have repeated the exercise in their own countries. Their results cannot be compared with each other: the measure, the groups compared and the end date change from one study to the next, and only the French study counts hires.

• United States, Stanford: employees aged 22-25 in the 2 most exposed occupation groups, compared with the 3 others, to June 2026. -11% compared with +10%, a gap that has been widening since the summer of 2025.
• United Kingdom, King's College London: junior positions (by seniority) in companies whose employees are highly exposed, compared with other companies, from 2021 to 2025. -5.8%, according to the university's press release on a working paper.
• Sweden, Örebro University: employees aged 22-25 in highly exposed occupations, compared with other occupations at the same employer, to early 2025. -5.5%. The general decline in job postings began with the interest rate rise of April 2022, 7 months before ChatGPT.
• Norway, IZA: employees aged 21-30, most exposed occupation group compared with the least exposed, to February 2026. An estimated gap close to zero, with high uncertainty. Among 22-25 year-olds, a gap appears in spring 2025 and widens until February 2026, which the authors regard as an emerging trend rather than an established effect. Employment of young developers falls by about 18%.
• Denmark, universities of Chicago and Copenhagen: wages and hours of 25,000 employees in 11 exposed occupations, all ages combined, depending on whether or not they use an assistant. No effect larger than 2%.
• Finland, ETLA: young employees, using the Stanford method, to September 2025. No systematic effect of exposure.

In the 3 countries where the studies isolate developers, their hiring or the number of young developers is falling. In the United States and Norway, the number of older developers is rising; in France, the decline in hiring affects all ages. The gap between exposed occupations as a whole and the others appears only in some countries.

Other causes affected the same occupations at the same time: the rise in interest rates in 2022 and 2023, the correction in digital-sector hiring after the pandemic, remote working. The Norwegian authors conclude that these countries may have experienced different shocks, rather than different effects of AI.

In the 3 studies that measure hires and departures, the decline comes through hiring. Departures of young people have fallen in the United States and in France alike, and promotions of junior staff have barely increased in the adopting companies studied at Harvard.

The decline appeared as early as 2023. For the Harvard economists, this is earlier than the effects of automation are usually felt, and they see it as an anticipatory adjustment, which they say they cannot test: companies may be stopping hiring for positions they expect to automate. If they overestimate what AI will be able to do in the short term, they add, the decline in junior hiring may only be temporary.

Company announcements point in different directions. In June 2025, Amazon's chief executive wrote to staff↗ that AI would reduce Amazon's total corporate workforce over the next few years. In February 2026, IBM's chief human resources officer announced that the group was tripling its entry-level hiring, developers included, with roles redesigned to involve less routine coding and more work with clients, according to Fortune↗. No published data verifies these statements.

In France, the companies surveyed by Apec↗ , the French association for executive employment, plan to recruit 16% more managers and professionals with 6 to 10 years of experience in 2026, and 1% more of those with under 6 years, whose recruitment fell by 13% from 2023 to 2025. Apec attributes this to economic uncertainty, without linking it to AI.

France has one particular feature: apprenticeships. It had 1.017 million apprentices at the end of 2025, and there were 3.4 times as many in higher education as in 2019. Among 21-29 year-olds in the most exposed occupations, the share of apprentices rose from 6% to 10% between late 2019 and late 2023; the HEC and Banque de France authors consider that this may affect their series, without explaining the gap between occupations on its own. In 2025, hiring of higher-education apprentices fell by 7.7%, mainly because of budget and regulatory cuts, according to OFCE, a French economic research centre.

For 3 economists in Barcelona, work done early in a career is also valuable for what it teaches. In their model↗, published as a working paper in August 2026, automating junior tasks can reduce long-term output when the cost saving is small: fewer young people choose occupations where experience matters, and the number of experienced professionals falls. They test this empirically on robot automation in the United States. The HEC and Banque de France economists reach the same conclusion: in their words, entry-level positions "are also those where the skills of tomorrow's managers and leaders are acquired through practice" (our translation). The Harvard economists give debugging code and reviewing legal documents as examples of such AI-exposed tasks.

A precedent has been observed outside generative AI. Researcher Matt Beane spent more than 2 years in operating theatres where robotic surgery was being introduced. With the robot, the senior surgeon regularly took back the controls, and residents got markedly less practice than in traditional surgery. Some learned nonetheless, outside the official training routes, through what Beane calls shadow learning.

The few direct measurements point the same way. In a Turkish high school, an experiment covered nearly 1,000 pupils in mathematics, some of whom worked with an assistant similar to ChatGPT. Their exercise results improved by 48%; once the tool was removed, their marks were 17% lower than those of pupils who had never had it. With a version designed to make the pupils do the work, the loss disappeared, with no measurable gain. In January 2026, Anthropic↗, the company that sells the Claude assistant, had 52 engineers, mostly junior, learn a programming library.

The largest gap concerned debugging, that is, the ability to check code. The time saved by the engineers with the assistant, about 2 minutes, was not significant.

The most serious objection comes from a frequently cited experiment. In a software company's support centre↗, researchers from Stanford and MIT tracked more than 5,000 agents. Those who had the assistant from the start resolved as many requests per hour after 2 months as colleagues without it did after 8 to 10 months. If AI saves months of training, the training question might resolve itself. The authors note that a company could then hire either more or fewer entry-level workers, and that their data do not show which.

They also measured what remains without the tool, when it was down. The equipped agents stayed above their pre-AI level, the more so the more closely they had followed the tool, and the authors see this as evidence of lasting learning. From their tenure curves, however, they say they cannot distinguish an agent who is learning from one who is following the recommendations.

So there are 3 measurements of what remains without the tool, and they diverge. At the support centre, the gain persists. Among the Turkish pupils and the Anthropic engineers, in 2 randomised experiments, skills decline when AI has done the exercise.

Our reading is that the difference comes down to the nature of the task. At the support centre, the tool suggested the best agents' answers to recurring requests, and following them amounted to learning them. For the pupils and the engineers, the tool gave the solution to exercises meant to build understanding. AI seems to help entry-level workers who are still doing the work, and to harm them when it does the work for them. No study yet measures this difference among entry-level employees in companies, or follows them over several years. For an organisation, the useful question is which tasks its entry-level staff still do themselves, and how much time its experienced employees spend reviewing their work.