A hiring signal worth taking seriously
The August 2026 revision of Stanford’s Canaries in the Coal Mine paper follows ADP payroll data through June. It reports that employment of workers aged 22–25 in AI-exposed occupations stands 19% below the path implied by less-exposed peers, with the divergence operating mainly through reduced hiring. The authors do not find comparable evidence of widespread displacement across the economy.
That is a relative employment gap in a particular dataset and comparison. It is not a count of jobs proven to have been eliminated by AI. Age, occupation, employer composition, other economic shocks, and the construction of the comparison matter when interpreting the result. The finding warrants sustained monitoring; causal attribution requires additional evidence.
Sources: Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Stanford Digital Economy Lab
A different result in Denmark
Humlum and Vestergaard link surveys of workers in exposed occupations to Danish administrative employment records. Their working paper finds substantial adoption and changes in tasks, but little early effect on recorded earnings and hours. Denmark’s institutions, occupations, observation window and measured outcomes differ from the Stanford study.
These results can coexist. A change in recruitment of young workers need not immediately change the average earnings of incumbent workers. Nor does a small average effect exclude losses in a particular group. A useful synthesis asks which workers, which outcome and which comparison produced the estimate.
Sources: Anders Humlum and Emilie Vestergaard, Becker Friedman Institute working paper
Job postings are an early indicator with limits
A New York Fed analysis finds that weaker postings in more exposed occupations do not display a clear new national break attributable to generative AI. Some divergence predates ChatGPT. A Dallas Fed analysis using a measure focused on automatable tasks finds weaker demand for more exposed positions. Its analysis compares occupations within industries and reports posting differences, not verified layoffs.
The disagreement is informative: exposure measures are not interchangeable. A measure of where AI could assist work may behave differently from a measure of tasks likely to be automated. Postings also move with interest rates, sector cycles and employers’ recruiting practices. Neither series should be relabeled as completed hires.
Sources: Federal Reserve Bank of New York, Liberty Street EconomicsFederal Reserve Bank of Dallas
Exposure does not predict a layoff count
The ILO’s 2025 refined index estimates that one in four workers globally is in an occupation with some generative-AI exposure. Its task analysis points primarily toward changes within jobs. Exposure describes overlap between capabilities and work, not whether employers will adopt a system or how demand will respond.
A credible employment claim needs the links between capability, adoption, work reorganization and labor outcomes. Forecasts often jump across those links. They can be useful scenarios, but they are not observations.
Sources: Paweł Gmyrek and colleagues, International Labour Organization
The distribution of gains is still unresolved
The published Generative AI at Work study finds a 15% average productivity gain in the customer-support setting it studies, with larger gains for less experienced workers. Its published estimate should take precedence over the different headline estimate in an earlier working-paper version. This is evidence about performance in a defined workplace, not proof of a nationwide wage gain or an economy-wide staffing reduction.
What would change the assessment
Opening’s assessment is that early-career hiring deserves particular attention and that broad claims of demonstrated economy-wide AI job destruction exceed the evidence reviewed here. This assessment would change with repeated results across independent datasets, credible comparisons around actual adoption, and consistent effects on hiring, separations, pay or hours.
The worker’s side of the story also requires evidence about workload, discretion and bargaining power. Rising output can coexist with worsening working conditions. Reporting should make those outcomes visible even when total employment barely moves.
Scope and limits
Selective review of studies relevant to current claims, not a systematic literature review or meta-analysis. Working papers can change. The included studies differ in country, sample, period and identification strategy; estimates must not be averaged together.
Sources
- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Stanford Digital Economy Lab. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Revised August 12, 2026; data through June 2026.
- Anders Humlum and Emilie Vestergaard, Becker Friedman Institute working paper. Large Language Models, Small Labor Market Effects. 2025 working paper; linked Danish survey and administrative data.
- Federal Reserve Bank of New York, Liberty Street Economics. Do Job Postings Show Early Labor-Market Effects of AI?. May 14, 2026.
- Federal Reserve Bank of Dallas. Job postings show early signs of AI automation impact. September 1, 2026.
- Paweł Gmyrek and colleagues, International Labour Organization. Generative AI and Jobs: A Refined Global Index of Occupational Exposure. May 20, 2025.
- Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, Quarterly Journal of Economics 140(2), 889–942. Generative AI at Work. 2025 published article.
Cite this article
Opening Research. “AI is changing work. The employment evidence is more specific than the headlines.” September 9, 2026. https://opening.works/research/ai-labor-market-evidence. Include your access date. Cite the original study when using its estimates.