Stanford economists say the labor market is quietly closing its front door on young workers exposed to AI. In an August 2026 update to "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," researchers led by Erik Brynjolfsson report that employment for workers aged 22 to 25 in the most AI-exposed occupations now runs 19% below their peers in AI-resistant jobs — a gap that measured 13% in last year's edition of the same paper.
The overall economy still looks fine. Across all age groups, the study finds little difference in relative employment between the most and least AI-impacted occupations. The damage is concentrated at the entry level: since 2022, employment for 22-to-25-year-olds in the top 40% of AI-impacted jobs has fallen roughly 11%, while employment for the same age group in the bottom 60% of jobs — those least touched by AI — has grown about 10% over the same period.
The Stanford team built its exposure ratings on a subsample of anonymized, high-frequency payroll data from ADP, cross-referenced with a labor-market impact metric from earlier academic work and the Anthropic Economic Index, which tracks how workers in various occupations actually use Claude. Google published a comparable index last month based on Gemini usage. The convergence of the payroll signal and the model-usage signal is what gives the 19% figure its weight.
Key facts
- 01Employment for workers aged 22 to 25 in the most AI-exposed occupations is now 19% below peers in less-exposed fields, up from a 13% gap last year.
- 02Since 2022, entry-level employment in the top 40% of AI-impacted jobs has fallen about 11%, while the least-impacted 60% grew 10%.
- 03The Stanford team used anonymized ADP payroll data and the Anthropic Economic Index to rate each occupation's exposure to Claude and similar models.
- 04Effects show up mainly as slower hiring, not layoffs or pay cuts, and are concentrated in jobs Anthropic labels 'automative' rather than 'augmentative'.
- 05Higher education dampens the effect: occupations with more college graduates show muted gaps between AI-exposed and AI-resistant work.
The mechanism is a hiring slowdown, not a firing spree. The researchers find the entry-level employment effects are driven by lower hiring rates in AI-impacted fields rather than by increased separations, and they show up almost entirely in headcount rather than in wages. Jobs are not opening; the ones that exist still pay about the same.
The distinction between automation and augmentation matters. Anthropic's index separates queries where Claude fully replaces a human task from queries where it makes a human worker more effective. Accountants, auditors, receptionists and information clerks skew heavily toward automative use. Chief executives and registered nurses skew augmentative. The occupations bleeding entry-level jobs are the automative ones.
“The findings are consistent with automation-oriented uses of AI substituting for labor while complementary uses are associated with flat or rising employment.”— Erik Brynjolfsson, Stanford economist and lead researcher
The paper also tests whether formal, codified knowledge — the kind taught in textbooks and standardized procedures — leaves workers more exposed than tacit knowledge acquired through practice and mentorship. Using O*NET's occupational database as a proxy, the researchers find that occupations heavier in codified knowledge show slower entry-level employment growth, while jobs richer in tacit knowledge show faster employment growth for mid-career and senior workers. Experience is holding its value; the textbook is not.
Higher education still helps. Occupations with a larger share of college graduates show much smaller gaps between AI-exposed and AI-resistant work. In jobs with few college graduates, the least AI-exposed occupations are growing while the most exposed are shrinking — the split is sharper the further down the credential ladder you go.
Brynjolfsson framed the risk directly in an interview with The Washington Post, arguing the trend describes a labor market that maintains its aggregate employment while cutting off the entry point for the next cohort. That is a distinct problem from a headline unemployment rate: the total number of jobs can look healthy while the pipeline for 22-year-olds narrows year over year.
There are caveats. The paper measures the first few years of a technology that is still changing quickly, and the exposure ratings depend on how a small set of AI products are being used right now — Claude and, by extension, Gemini. If augmentative use cases grow faster than automative ones, or if firms rediscover the value of training junior staff on AI-assisted work, the widening gap could stall. The paper's own finding that augmentative jobs show flat or rising employment leaves that door open.
For the AI industry, the Stanford numbers are the clearest empirical signal yet that model deployment is showing up in payroll data, not just productivity slide decks — and that the impact is landing on the workers who cost the least and generate the least revenue. That is a politically combustible outcome for a sector that has spent two years arguing AI would augment rather than replace. Expect it to sharpen the case for regulation aimed specifically at entry-level hiring disclosures, and expect the labs whose usage data underpins these studies — Anthropic and Google in particular — to face pressure to publish more of it, not less.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
