Stanford's Updated Canaries Study: AI's Entry-Level Jobs Gap Widens to 19%

For two years, the debate over AI and jobs has alternated between two poles: a coming apocalypse that wipes out whole professions, or a productivity boom that leaves everyone better off. Stanford’s Digital Economy Lab has just published the strongest evidence yet that the reality is more precise — and more uncomfortable — than either camp predicted.

On August 12, 2026, the lab released a revised edition of “Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence,” by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen. The update extends the team’s analysis of anonymized ADP payroll data through mid-2026, and its headline number has moved in the wrong direction for young workers: employment among 22-to-25-year-olds in highly AI-exposed occupations now stands about 19% below where it would be had it kept pace with their peers in less-exposed occupations. When the paper first appeared in August 2025, that gap was roughly 13%; at the July 2025 data vintage it measured 15%. The divergence has widened steadily ever since.

What the study actually found

The researchers built their analysis on a large subsample of high-frequency payroll records aggregated by ADP, rating each occupation’s exposure to AI using established academic exposure measures (Eloundou et al., 2024) alongside the Anthropic Economic Index, which tracks how occupations actually use Claude in day-to-day work.

Their six facts, updated with the new data:

1. No economy-wide displacement. Across all ages, there is little relative employment difference between the most and least AI-exposed jobs. The aggregate labor market is not collapsing.
2. But young exposed workers are falling behind. The 19% gap for 22-25-year-olds is the study’s sharpest signal. In levels, employment in the two most-exposed quintiles fell about 11% between November 2022 and June 2026, while the same age group in the three least-exposed quintiles grew about 10%.
3. Experienced workers show no comparable gap. Older cohorts in the same exposed occupations are holding steady — a clean age gradient that older explanations of labor-market churn don’t predict.
4. The mechanism is hiring, not firing. The adjustment operates primarily through reduced hiring of young workers, not increased separations. Companies aren’t replacing junior staff with AI so much as quietly not onboarding them in the first place.
5. Automation beats complementation. Declines concentrate in occupations where AI usage is automating rather than complementary. Where AI augments experienced workers, employment is flat or rising.
6. Wages haven’t moved yet. So far the adjustment shows up in employment rather than base pay.

The codified vs. tacit knowledge split

The most intellectually interesting addition in this revision is a distinction between two kinds of knowledge. Employment has declined among young workers in occupations that run on codified knowledge — formal, standardized, documented knowledge teachable through textbooks and written procedures. Conversely, employment has increased among experienced workers in occupations that depend on tacit knowledge, the kind built through practice, mentorship, and repeated exposure to real situations.

The interpretation is intuitive: generative AI is exceptionally good at reproducing and applying knowledge already encoded in text and digital information. Entry-level work has historically been exactly that — applying what you just learned from documentation. The junior tasks AI absorbs first are the ones that live in manuals. The senior judgment that comes from a decade of edge cases remains stubbornly hard to replicate — and, for now, increasingly valuable.

The paper also notes that women face greater AI exposure on average, a heterogeneity the team says it will monitor as the data evolve.

Correlation, causation, and honest caveats

The authors are unusually careful about what the data can and cannot say. These are descriptive patterns, not causal estimates. Several alternative explanations were tested and found insufficient: the divergence survives excluding tech firms and computer occupations, controlling for interest-rate exposure and remote work, and swapping in alternative AI-exposure measures. The gap also kept widening through mid-2026, well after interest rates peaked, and by November 2022 exposed young workers had already recovered to pre-pandemic levels — meaning the decline pushes below a recovered baseline rather than merely unwinding a pandemic distortion.

But the caveats are real too. Gaps shrink when accounting for education. Some differential trends predate generative AI. And the estimated gaps are larger in the ADP sample than in national survey benchmarks, with discrepancies concentrated in education, health care, and public administration. The authors explicitly refuse to treat their own paper as definitive, arguing that cumulative evidence across studies will matter more than any single result.

A dashboard, not a verdict

Perhaps the most consequential move is infrastructural. The Stanford Digital Economy Lab has launched AI Economic Indicators, a set of high-frequency measures of AI’s economic effects, and the accompanying Canaries Dashboard will re-run the paper’s key results every month. The goal, as the lab puts it, is “not to declare the labor-market effects of AI settled” but “to make them measurable.”

That framing matters for how the industry should read this study. The 19% figure is not a prediction of permanent decline for an entire generation — it is an early indicator, designed to be tracked continuously. If generative AI is reshaping opportunities for specific groups of workers before aggregate statistics show anything, the point is to catch it early, understand the mechanisms, and watch whether it spreads to older cohorts or reverses as complementary roles mature.

Why it matters

The study lands amid mounting anecdotal unease about the junior rung of the career ladder — slowed new-grad hiring in software, customer service, and content roles — and gives it a rigorous empirical backbone. For employers, it suggests the “AI does the grunt work” model has a hidden cost: if entry-level positions are where humans acquire the tacit knowledge that AI cannot yet replicate, cutting them off may eventually hollow out the pipeline that produces experienced workers in the first place. For policymakers and educators, the age gradient — young workers hit, older workers untouched — is precisely the pattern that standard retraining rhetoric is worst at addressing. And for the AI industry, it is a reminder that labor-market effects are arriving unevenly, quietly, and through the hiring channel, long before any headline “job losses” number confirms them.

The canary metaphor is apt in one more sense. Coal miners carried canaries not because the birds’ distress proved disaster, but because it bought time to act. A widening, measurable, age-specific employment gap is that kind of signal — and as of this revision, it is getting louder.