Brynjolfsson, Chandar & Chen (2025) — Canaries in the Coal Mine?
TwinLadder Research Brief · Source Summary · May 2026
Companion reference to The Authority Gap.
Why this paper matters
For nearly three years after the public release of ChatGPT in November 2022, the debate about AI's impact on jobs ran on speculation. Forecasters argued, founders made claims, surveys produced contradictory readings. The empirical record was thin because most of the data sources good enough to detect a labour-market shift — administrative payroll, monthly earnings, individual-level employment — are not publicly accessible at scale.
In November 2025, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, working through Stanford's Digital Economy Lab and using high-frequency administrative data from ADP — the largest payroll provider in the United States, covering millions of workers across tens of thousands of firms — provided the first large-scale empirical evidence that generative AI is already reshaping the labour market. The shift is concentrated, asymmetric, and visible primarily in early-career workers.
This is a paper widely cited and frequently misread. This brief sets out what the six "facts" the authors document actually say, what they do not say, and how they connect to the broader argument about authority migration under AI.
What the data are
The study uses monthly, individual-level payroll records from ADP through September 2025, covering millions of workers. The panel structure lets the authors track individual employment over time. The granularity — monthly observations of named workers, age, industry, occupation, firm — is rare in this literature; most prior labour-market studies of AI rely on surveys or annual aggregates, both of which are too noisy to detect short-run shifts.
AI exposure is measured at the occupation level, using existing taxonomies of which jobs are most exposed to generative AI tools. The authors then examine whether employment trends differ across exposure quintiles, age cohorts, and over time.
The six facts
The authors enumerate six findings that emerge from the data. Each is deliberately framed as a "fact" — an empirical observation the data support — rather than as an interpretation.
Fact 1. Substantial employment declines for early-career workers in AI-exposed occupations. Workers aged 22–25 in the most AI-exposed occupations show large declines in employment. Workers in less-exposed occupations, and more-experienced workers in the same occupations, show stable or growing employment. The asymmetry is the central finding.
Fact 2. The asymmetry is concentrated, not aggregate. Overall US employment continues to grow. But employment growth for young workers in AI-exposed occupations has been stagnant since late 2022. From late 2022 to September 2025, workers aged 22–25 in the most AI-exposed occupations saw a 6% decline in employment, compared to a 6–9% increase for older workers in the same occupations.
Fact 3. Automation, not augmentation, drives the decline. The authors distinguish empirically between AI uses that automate work and AI uses that augment it, using observed query patterns to Anthropic's Claude as a measure of substitution versus complementarity. Employment declines are concentrated in occupations where AI primarily automates. Occupations in which AI use is most augmentative show employment growth, even for young workers.
Fact 4. The findings survive controls for firm-time effects. A standard objection to age-occupation correlations of this kind is that they could be driven by industry- or firm-level shocks (an interest-rate cycle, a sector downturn) that happen to correlate with AI exposure. The authors run an event-study regression with firm-time fixed effects and find a 15-log-point relative employment decline for the most-exposed quintile compared to the least-exposed quintile, for workers aged 22–25 specifically. Other age groups show no statistically significant effect. This is the result that converts the correlation into something closer to a causal claim.
Fact 5. The adjustment runs through employment, not wages. Annual salary trends differ very little across age and exposure quintile. The labour market is adjusting to AI by hiring fewer people in the affected categories rather than by reducing pay. The authors note that this is consistent with downward wage stickiness and that AI's wage effects, if any, may take longer to materialise than its employment effects.
Fact 6. The findings are robust to a range of alternative explanations. The patterns are not driven by computer occupations alone (excluding tech firms and remote-eligible occupations does not eliminate the result). The AI exposure taxonomy did not predict employment outcomes for young workers in earlier periods, including during the COVID-19 unemployment spike — meaning the effect appears specifically when generative AI becomes widely available, not before. The pattern holds for both college-educated and non-college occupations.
The headline figure most often cited from the paper — a roughly 16% relative employment decline for 22–25-year-olds in the most AI-exposed occupations after controlling for firm-level shocks — is the consolidated form of Facts 1, 2, and 4 read together.
What the paper does not show
Three things this paper does not establish, and that responsible citation should not impute to it:
- It does not show AI is causing the decline. It shows the decline is concentrated where AI exposure is highest, after plausible controls. Causal interpretation is highly probable but not proved; the authors are appropriately cautious.
- It does not say what the displaced young workers are doing instead. The data are about employment in AI-exposed occupations. They do not track whether 22–25-year-olds who would have entered software development are now entering nursing, plumbing, or unemployment. Other research will have to answer that.
- It does not generalise outside the United States. The data are ADP payroll, which is US-only. The labour-market institutions, hiring norms, and AI deployment patterns of European and other markets may produce different effects. Early evidence from Europe is consistent with the US pattern, but is not established at the scale of this paper.
How this connects to the Authority Gap
The Authority Gap research piece argues that authority over AI-shaped decisions is moving across the org chart while legal accountability stays put. One direction of that movement is off the org chart entirely — the entry-level work that used to populate the bottom of the pyramid is shrinking. The pyramid that used to feed mid-career and senior expertise is being eroded from the bottom.
Brynjolfsson, Chandar, and Chen's data are the empirical anchor for that observation. The structural shift they document is not a forecast; it is already in payroll. The succession-planning implication — that the senior roles your organisation will need to fill in five years are produced by entry-level pipelines that AI is currently contracting — is the bridge from this paper to the HR-specific Authority Gap brochure, which argues that this is fundamentally a People function problem the People function has not yet been told it owns.
Citation
Brynjolfsson, E., Chandar, B., & Chen, R. (13 November 2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab Working Paper. digitaleconomy.stanford.edu
Working paper PDF: Canaries_BrynjolfssonChandarChen.pdf
A primer by one of the co-authors is available at bharatchandar.substack.com.
TwinLadder Research Briefs are short reference summaries of the foundational sources cited in our research pieces. They are not commentary; they are background reading. Companion to the Authority Gap launch series, May 2026.