Stanford Payroll Data Ties AI to a Shrinking Entry-Level Job Market

A working paper from Stanford’s Digital Economy Lab, built on ADP payroll records covering millions of U.S. workers through June 2026, finds no sign of AI causing broad, economy-wide job losses — but a specific, widening gap for the youngest workers. Economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen report that employment for 22-to-25-year-olds in AI-exposed occupations is now about 19% below where it would sit had it tracked less-exposed peers, up from 15% at the July 2025 data vintage. In levels: headcount for that age group in the two most AI-exposed quintiles fell roughly 11% between November 2022 and June 2026, while the same age group in the three least-exposed quintiles grew about 10%. Every age group over 25 grew over that period — by as much as 11% for workers 35 to 40.

The authors trace the decline to hiring, not firing. Separation rates for young workers actually fell in AI-exposed occupations, at least as much as in less-exposed ones — the opposite of what a displacement story would predict. The gap instead opened because firms hired fewer entry-level workers into exposed roles. An occupation-level regression across exposure quintiles puts the decline for 22-to-25-year-olds in the most-exposed quintile at about 18 percentage points relative to the least-exposed quintile. That’s a more conservative figure than the firm-level, fixed-effects regression the same authors headlined in earlier drafts of this paper (a 13% relative decline in July 2025 data, 16% in September 2025) — an estimate they’ve since de-emphasized in favor of the simpler quintile comparison, noting it became more sensitive to data-pipeline changes than the headline number.

The effect also splits along how AI is used. Using both a GPT-4-based exposure score (Eloundou et al., 2024) and the Anthropic Economic Index’s classification of Claude queries as “automative” or “augmentative,” the authors find the decline concentrated specifically where AI substitutes for tasks: each standard deviation of automation exposure is associated with a statistically significant ~10% additional employment decline for 22-to-25-year-olds, while “augmentative” exposure shows no comparable effect for the young and a positive, significant association with employment for workers 41 and older. They tie this to a codified-versus-tacit-knowledge split — occupations built on formal, textbook-taught skills see slower entry-level growth, while occupations built more on hands-on experience see faster growth for older workers.

The pattern survives excluding the tech sector and computer occupations, controlling for interest-rate exposure and remote-work amenability, across five alternative AI-exposure measures, and separately for part-time and temporary workers and for men and women (women, who work in more AI-exposed occupations on average, show somewhat steeper declines). It holds up less cleanly against education: controlling for an occupation’s share of college graduates cuts the estimated decline roughly in half, from 18 percentage points to 9 — the one control the authors flag as producing “the greatest attenuation,” and in one sample specification the residual effect loses statistical significance entirely. Base compensation, meanwhile, shows little divergence between exposed and non-exposed young workers, so whatever adjustment is happening shows up in headcount rather than pay. The authors describe all of this as descriptive rather than causal, and note the magnitude is larger in their ADP sample than in Census survey data, even where the direction agrees.