Between November 2022 and June 2026, employment of 22-to-25-year-olds in the two most AI-exposed groups of occupations fell about 11%. Over the same stretch, employment of that age group in the three least-exposed groups grew about 10%. Those figures come from the Stanford Digital Economy Lab, which published a revised version of its “Canaries in the Coal Mine?” paper on August 12, 2026.
The authors, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, drew on ADP payroll data. They say the paper traces how employment has moved since ChatGPT was released, split by age and by how exposed an occupation is to AI.
A 19% gap for young workers
The headline number: employment among workers aged 22 to 25 in highly exposed occupations now sits about 19% below where it would be had it kept pace with similarly aged workers in less-exposed jobs. Experienced workers show no comparable gap.
The shortfall has grown. It was 15% in the July 2025 data and reached 19% as of June 2026. The lab says the adjustment appears to run mainly through reduced hiring of young workers, not through more people leaving jobs. Pay has barely moved so far. The change shows up in headcount, not in base wages.
The lab does not report widespread, economy-wide job losses tied to AI.
Automation versus assistance
The declines cluster in occupations where AI tends to automate human tasks. Where it works more as a complement, employment is flat or rising, especially among experienced staff.
The revision adds a new angle. Young workers lost ground in occupations that lean on codified knowledge, meaning the formal, documented kind taught through textbooks and written procedures. Experienced workers gained in occupations that depend more on tacit knowledge, picked up through practice and mentorship. The authors say that fits the idea that generative AI is good at reproducing what has already been written down. They also found that women face greater AI exposure on average.
What the authors caution
The researchers are blunt about the limits. “These are descriptive patterns, not causal estimates of the effect of AI,” they write. The data alone cannot show how much of the divergence generative AI caused rather than other forces in the labor market.
They tested some rival explanations. The gap persists when technology firms and computer occupations are excluded, when interest-rate exposure and remote work are controlled for, and when other measures of AI exposure are used. It also kept widening through mid-2026, well after interest rates peaked.
But the gaps shrink once education is taken into account, some differing trends predate widespread generative AI, and the estimated gaps are larger in the ADP sample than in national survey benchmarks. “We do not view this paper, or any single study, as definitive evidence of AI’s labor-market effects,” the lab wrote. Its Canaries Dashboard, built with ADP Research, will refresh the paper’s key results every month.
