Research check · 23 October 2026
Young workers’ employment fell again in AI-exposed U.S. jobs. The cause is still open.
The Stanford–ADP September payroll update shows a sharper year-over-year fall among 22–30-year-olds in occupations rated highly exposed to AI. It counts workers on payroll in a selected U.S. firm panel, not layoffs caused by AI.
What changed
Employment in high-exposure occupations was 0.5% lower than a year earlier in September 2026; the least-exposed group grew 1.0%. Among workers aged 22–25, the split was −3.7% versus +0.2%. For ages 26–30 it was −2.8% versus +0.9%. ADP reports 36 straight months of year-over-year contraction for the 22–25 high-exposure group. These percentages are changes within groups, not percentages of all workers losing jobs.
This is a fresh month of the existing Canaries tracker, not a new independent study. Stanford’s methodology uses firms with five years of ADP payroll history and workers mapped to occupations. It excludes firm entry and exit, is not nationally representative, and groups jobs by possible AI exposure rather than measured employer use.
What the signal cannot prove
The latest data strengthen the warning that young workers in some exposed occupations face a weaker employment path. They do not show how many jobs AI removed. A four-country hiring study finds that remote-work exposure weakens the apparent GenAI link to lower junior hiring shares. It uses a different sample and stops in 2025. French administrative hiring data show a later entry gap, but also cannot isolate AI. Broad U.S. tracking still finds no clear economy-wide AI footprint; it tests a different outcome and scale.
German counterpoint
German IAB administrative records add a counterpoint: full-time employment among 20–29-year-olds fell 3.8% in more AI-exposed occupations versus 2.6% in less-exposed occupations from 2022 to 2025. After accounting for remote-work suitability and cyclical vulnerability, the AI-exposure association is negative but not statistically significant; the business-cycle association is stronger. Germany’s full-time stock, age band and earlier period differ from the U.S. panel and French hiring flows. This is not a direct replication test.
The next check
Track the same employers’ actual AI and remote-work adoption, hiring, departures and headcount by age. Until then, a personal job-loss prediction from these group averages would be unwarranted.