Stanford’s August 2026 canaries: 19% is a relative gap, not jobs lost to AI

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “Canaries in the Coal Mine?” August 2026 revision (authors’ clear explanation). This is the up-to-date revision of Gates’s linked 2025 paper, not its original numeric result. Read 30 September; part of employment synthesis.

Design and result

Monthly ADP payroll, June 2026 endpoint; analysis uses a balanced panel of surviving ADP client firms with 3.5–5 million full-time employees per month, job-title occupation crosswalks with some missing occupation imputed; AI occupational exposure based on GPT-4 task ratings and, separately, Claude query-based automation/complement classifications. The employment of 22–25-year-olds in the two most exposed quintiles fell roughly 11% between November 2022 and June 2026, while in three less-exposed quintiles it rose roughly 10%. The authors’ 19% shortfall is what exposed young-worker employment would lack had it kept pace with less-exposed young-worker employment; it is not a measured 19% AI-caused fall, a 19% unemployment rate, or economy-wide job loss. Experienced workers show no analogous exposure gap; reduced young hiring rather than extra separation drives the pattern. AI-automation-classified jobs have the stronger decline; complementary-use jobs do not. Base pay (not variable compensation) changed much less.

Fair reading and objections

The authors say explicitly these are descriptive facts, not causal estimates; no economy-wide displacement observed. Stronger than a tech-layoffs anecdote: divergence persists omitting tech/computer jobs, accounting for remote work and differential rate exposure, and with firms entering or leaving the sample; timing after rate peak, and an independent Census industry-state study lends support. But pandemic pretrends exist; controlling occupation education attenuates gaps; adding firm-time controls gives directionally consistent yet more specification-sensitive estimates after pipeline revisions; ADP panel overrepresents large firms, manufacturing, wholesale and AI-exposed occupations and differs from survey-based national estimates in education/health/public sectors. Nor do measured occupational exposure and model query mix reveal the treated firm’s actual AI use. One observational sample cannot determine displacement of all young workers versus occupational reallocation or lost entry rungs.

What would discriminate

Track actual AI adopters and matched nonadopters with hires, separations, firm entry, output, prices, workers switching firms/occupations, and long pre-trends; estimate how many young workers remain employed anywhere and at what wages. See independent Denmark and multinational adopter comparisons and Lee Tucker’s 2026 Census study, whose administrative industry-state panel finds a regression-adjusted 12% relative decline for ages 22–24 over ten quarters but itself reports COVID-era pretrends and does not prove attribution. Avoid treating Tucker’s different age/denominator as a simple replication of 19%.