Stanford Updated Its "Canaries in the Coal Mine" Paper With ADP Payroll Data: Employment for 22–25-Year-Olds in AI-Exposed Jobs Sits 19% Lower

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen of the Stanford Digital Economy Lab revised "Canaries in the Coal Mine?" in August, now drawing on ADP payroll records covering millions of US workers through June 2026. The sharpest of its six facts: employment for 22-to-25-year-olds in AI-exposed occupations runs about 19% below the less-exposed comparison group, with no equivalent gap for experienced workers — and that gap has widened steadily since it was first documented in August 2025 at around 13%. The mechanism is reduced hiring rather than layoffs; declines concentrate where AI substitutes for human tasks while employment holds or rises where AI complements workers; and the adjustment shows up in headcount, not base pay. The authors stress these are descriptive early indicators, not causal estimates.

Look at the Data Source Before the Conclusion

Data is where this genre of research usually goes wrong. This one doesn't use surveys or scraped job listings; it uses ADP's administrative payroll records — data generated when people actually get paid — covering millions of US workers and updated monthly through June 2026. The same source feeds the lab's Canaries dashboard, built with ADP Research, spanning roughly 4.6 million workers across more than 730 occupations. Because the unit is payroll, the "reduced hiring" claim holds up: new hire records and separation records are distinct events in the data, and what the authors observe is fewer of the former. Companies aren't cutting young workers so much as opening the door less far.

"No Collapse" and "Harder to Get In" Are Both True

The first of the six facts is that there is no evidence of widespread, economy-wide displacement. That one tends to get dropped in the retelling, and it matters — Brynjolfsson himself has said an AI "job apocalypse" looks unlikely. What is happening is distributional: concentrated in ages 22–25, and in occupations where AI substitutes for the work, while occupations where AI complements people are flat or growing, especially for experienced staff. Pay has barely moved; the adjustment runs through headcount. Don't skip the stated limitations either. These are descriptive indicators rather than causal estimates, the effects are more pronounced in the ADP sample than in national survey benchmarks, and some of the divergence predates generative AI. The 19% is a relative gap, not "19% of jobs disappeared."

What Transfers and What Doesn't

The data is American, and copying the conclusion onto another labor market will produce errors — industry mix, graduate volumes and hiring cycles all differ. The structural claim is what travels: if entry-level roles happen to consist of exactly the tasks tools eat first (boilerplate code, tier-one support, basic data cleanup), then the "spend two years on grunt work, then step up" path gets shorter or disappears. For people entering the market, the valuable part shifts from completing a task to judging whether it was done right — reviewing model output, localizing failures, turning vague requirements into verifiable steps, all things models still do unreliably. For teams hiring, cutting junior headcount is the cheapest choice this quarter, but a team with no mid-level engineers three years out has only deferred the training cost, not avoided it.

via: The Stanford Digital Economy Lab paper page, "Canaries in the Coal Mine?", the Canaries dashboard, Ars Technica