Early-Career Employment Decline in AI-Exposed Occupations

the gap between unemployment among 20-24-year-olds and the overall rate is close to a multi-decade low

Since before ChatGPT, employment in the most AI-exposed occupations is down around 6% relative to the least-exposed occupations, with the gap reaching 19% among workers aged 22–25

The clearest available evidence that the early-career problem is a credential-specific problem rather than an age-specific one. Ranking each group's 12-month moving-average unemployment rate against its own history back to 2003, Levanon finds young college graduates aged 22-34 sitting at the 70th percentile of their own history while young workers of the same age without a degree sit at the 25th. The two series moved together for two decades and decoupled in 2024. Young graduates are roughly 20 million workers, about one in eight in the labor force. The figures are percentile ranks against each group's own history rather than an employment decline against least-exposed peers, and the article makes no AI attribution for the gap, so it is recorded as an overlay.

"The jobs at most risk are entry- and mid-level, and the new jobs being created will mostly require skills that take many years to learn." Gates adds: "I'm especially worried about young people, who will enter a workforce with fewer entry-level openings." He cites the Stanford Canaries paper as his evidence that "employment fell significantly among young workers in jobs that are especially vulnerable to replacement, but not among their older colleagues." Commentary essay; no original data.

"Quantitatively, the effects mean that an industry for which the junior bias was one standard deviation stronger saw an additional decline of about 10% in employment." US Census and ACS microdata 1970-2020 at the 1990 Census industry classification, using occupation-level robot-exposure measures from Garg et al. (2026); regressions control for year and industry fixed effects and for overall industry robot exposure interacted with year. The same specification finds junior-biased robotization raised industry average age and steepened the wage-experience profile. Robot automation, not AI, and an industry-level rather than age-22-25 cut, so the magnitude is not directly comparable to the ADP and Revelio series plotted here.

"Goldman analyzed employment growth across more than 800 occupations and found that AI-related headwinds were the strongest among entry-level workers. It also found an additional, though smaller, negative effect among occupations considered to have a high risk of displacement from AI." The magnitude: "Across the broader labor market, a 10% occupational exposure to AI was associated with only a 0.1 percentage point drag on annual headcount growth in France, Canada and the U.S. But for entry-level workers, the impact ranged between more than 0.6 percentage point (Australia) and over 0.2 percentage point (U.S.)." This is a semi-elasticity of annual headcount growth per unit of occupational exposure, defined by seniority rather than age 22-25, and it carries no exposure gap that would convert it to the level shortfall this chart plots, so it is recorded as an overlay.

Employment of young workers (ages 22-25) in AI-exposed occupations now stands 19% below where it would be had it kept pace with that of their less-exposed peers; experienced workers show no comparable gap. Descriptive kept-pace measure on an ADP balanced payroll panel of 3.5-5 million workers per month, January 2021 through June 2026. Earlier vintages of this study headlined regression-adjusted estimates (13% at July 2025 data, 16% at September 2025 data); by the kept-pace measure the shortfall was 15% at the July 2025 vintage and has widened to 19% as of June 2026.

A direct check on whether labour-market entrants are the margin where AI is biting, and the answer here is no: "Job-finding rates for new entrants have declined in recent years, but the drop is modest relative to the declines among job losers and job leavers. Entrants do not appear to be the margin where the action is, and we set them aside in what follows." Using CPS reason-for-unemployment categories (job losers, job leavers, entrants and re-entrants), the brief finds the action instead among strongly attached incumbents in AI-exposed occupations. This measures job-finding probability rather than the employment shortfall of workers aged 22-25 against least-exposed peers, so it is recorded as an overlay - but it is a meaningful caution that an entrant-specific reading of the AI evidence may be misattributing an incumbent phenomenon.

"The report estimates that 6.1 per cent of jobs currently held by young people aged 15 to 29 fall in the categories of those most exposed to AI" - 55.8 million of 913.8 million youth jobs globally, using gradients 3 and 4 of the ILO occupational exposure scale (mainly clerical and support work). Under an illustrative scenario: "If only 10 per cent of the jobs that are exposed to AI disappear in full, that would translate into 5.6 million employed youth at the global level either facing unemployment, shifting into another job or exiting the labour force." Eastern Asia alone accounts for 1.5 million of those transitions, Latin America and the Caribbean 785,000, South-Eastern Asia and the Pacific 716,000, Northern America 618,000. ILO is explicit that "the degree to which risk translates to job loss is still debatable." A global youth-population exposure share under a hypothetical disappearance scenario, not a measured US employment decline for ages 22-25, so it is recorded as an overlay.

Employment for younger workers in the most AI-exposed occupations is down by 13% relative to the least exposed occupations, since pre-ChatGPT — a much larger decline than for older workers. Event-study with two-way occupation and month fixed effects, standard errors clustered by occupation, ages 22-25, built on online professional profiles rather than payroll records.

We do find some suggestive evidence that hiring rates for young workers in highly AI-exposed roles have weakened over the past year or so. [...] But this evidence for young worker displacement should be interpreted with caution. [...] from 2022 to now, the US experienced the largest non-recessionary labor market slowdown on record (the 'immaculate disinflation'). This coincided with a 'low hire, low fire' labor market. This kind of labor market hits early-career entrants hardest. Right now, young workers may be struggling to find jobs for macroeconomic reasons other than AI.

In the most-exposed quintile, employment among early-career women has been contracting at 4.5% per year since the introduction of ChatGPT in November 2022, while employment for men of the same age is contracting at 2.5% per year.

"Data from Australia, Canada, the European Union and the United States suggest that the role of LLMs in explaining the unemployment gap of young labour market entrants remains limited, both for those with and without a graduate degree." LinkedIn AI-hiring data date intensive firm LLM use to mid-2023 through early 2024, yet "no turning point in the unemployment gap trend for college graduates is visible around those dates... in any of the countries analysed." On postings: the ratio of new online postings in top-quintile versus bottom-quintile Language Model Exposure occupations (Felten, Raj and Seamans score, Lightcast data) "exhibits no break at the end of 2023," and Box 1.6 confirms this holds for US postings flagged junior by Lightcast's seniority classifier. OECD attributes observed declines instead to macro sensitivity: LLM-exposed occupations concentrate in information, finance and insurance, and professional and technical services, which cut hiring quickly when capital costs and uncertainty rise. This is an attribution finding, not a magnitude - OECD documents that the youth unemployment gap is real and widening, only that LLMs are unlikely to be its main cause - so it cannot be plotted as a level on this chart and is recorded as an overlay.

According to one Stanford study, in the occupations most susceptible to greater automation, such as software engineering or customer service, entry-level employment has already declined by 16 percent relative to the least-exposed occupations.

We therefore investigated effects for workers 34 and younger, finding mixed evidence of AI effects for this group... roughly half a percentage point increase in the entire sample, and more for the 16-34 year old subsample - but both statistically insignificant.

Regression-adjusted employment of early career workers in the most AI-exposed quintile of industry-state cells declined by 12% over the 10 quarters following the introduction of ChatGPT, even as employment in less-exposed industries rose.

They find that among 22-25 year olds employment in the top two quartiles of AI exposure fell about 12 percent relative to employment in the bottom quartile.

An event study documents an accelerating decline in employment of 22-25-year-olds in high-AI-exposure occupations, reaching 5.5 per cent by early 2025 relative to less exposed occupations within the same employers.

In het derde kwartaal van 2025 werkten ruim dertien procent minder jongeren in de meest vatbare beroepsgroepen dan in het vierde kwartaal van 2022, terwijl de werkgelegenheid in andere beroepsgroepen juist drie procent steeg.

Entry-level hiring rates have declined 23 percent compared to pre-pandemic levels, a steeper drop than the 18 percent decline for overall hiring.

Workers age 22-25 in most AI-exposed occupations experienced 13% employment decline since 2022, driven by fewer workforce entrants rather than higher separations. Employment share for AI-exposed occupations fell from 16.4% (Nov 2022) to 15.5% (Sep 2025).

Early-career workers (ages 22-25) in the most AI-exposed occupations have experienced a 16 percent relative decline in employment even after controlling for firm-level shocks. By September 2025, employment for software developers aged 22-25 declined nearly 20% compared to its peak in late 2022.

In jobs with high AI exposure, employment for 22- to 25-year-olds fell 6% between late 2022 and July 2025. Software developers saw a 20% early-career decline. Employment among workers 30 and older grew 6-13%.

"In Growth Roles, GenAI automates foundational work that historically served as the training ground for future expertise. The result: a narrowing of career entry points." "Our analysis reveals that approximately 1 in 8 US workers (17.8 million people) are currently in occupations where there could be considerably less entry-level opportunity as a result of GenAI." The countervailing finding: "About 1 in 5 workers (28.6 million people) are in fields where GenAI could take on technically sophisticated tasks, reducing hard-skills requirements and expanding access to well-paying jobs and opportunities." The report stresses that the narrowing side is concentrated in better-paying work: Growth Roles pay "28% more on average than Mastery Roles, a difference of $20,000 annually" and are "51% more likely to require a college degree," so "AI will disproportionately restrict entry points to high-value careers." Neither figure is a measured employment decline - both are counts of workers in affected occupations, projected forward - so they are carried as overlays rather than plotted against this graph's ages 22-25 decline metric.