Stanford's Institute for Economic Policy Research went looking for an AI employment shock in the national labor data and did not find one. Its policy brief reports that unemployment in the most AI-exposed occupations has not worsened faster than in the least exposed since 2022 - the pattern looks like a general labor-market softening, not a visible AI layoff wave. The brief's more uncomfortable finding is narrower: in two specific occupations, the youngest workers have been falling behind their older colleagues.
Key facts
- The SIEPR policy brief uses quarterly labor-force microdata from 2015 to 2026, splitting occupations into five AI-exposure bands.
- The most-exposed band's unemployment did not worsen more than the least-exposed band after 2022.
- The early-career warning covers customer service representatives and software developers, ages 22 to 25, in payroll data from a five-year balanced set of client firms.
- Stanford's own cited authors later found the divergence becomes notable in 2024, not immediately after ChatGPT's 2022 release.
The measurement everyone misreads
The headline unemployment chart does not measure whether a worker uses AI, or whether an employer automated a job. It assigns each occupation an exposure score, drawn primarily from the AI Occupational Exposure index built by Felten, Raj and Seamans. That index links progress in AI applications to specific workplace abilities catalogued in the O*NET occupational database, then aggregates by how important and prevalent those abilities are in each occupation.
The authors of that index are explicit about something almost every citation of it drops: exposure is agnostic about whether AI substitutes for a worker or complements one. It is a task-capability proximity map, not an automation-risk meter. A high score means AI is getting good at things this job involves. Whether that ends up replacing the worker or making them faster is exactly the question the index does not answer.
The extremes are intuitive once you know that. Text-and-analysis-heavy roles such as genetic counselors and financial examiners score high. Physical roles such as dancers and construction helpers score low. The aggregate result covers all coded occupations grouped into quintiles - not "AI jobs" versus everyone else.
The caveat Stanford puts in its own brief
This is the part that separates careful reading from a headline. The authors of the underlying early-career study later reapplied their analysis with stricter firm-by-time controls. Under that specification, the young-worker divergence becomes notable in 2024 rather than immediately around ChatGPT's release - and they explicitly name interest rates, pandemic-era overhiring, remote work, and other contemporaneous changes as complications for causal attribution.
That timing shift matters enormously. A decline that starts the month a product launches invites a causal story. A decline that becomes visible two years later, during a period of rate-driven hiring contraction across the whole economy, does not. It makes "AI has already replaced the junior workforce" an overclaim - without erasing the later, persistent relative decline, which is real and which nobody has explained away.
An independent directional check points the same way. A Census Bureau working paper using matched employer-employee administrative data finds reduced hiring and employment for 22-to-24-year-olds in the most AI-exposed industry-state cells after ChatGPT - and reports the same pre-existing pandemic-era trend shifts that limit what can be attributed to AI. Two studies converging is worth watching. It is not causal closure.
The brief also finds that firm adoption measures disagree wildly with each other, because Census business surveys, household surveys, a regional Fed survey and a corporate-spending index sample entirely different populations. Adoption is growing on all of them. Broad, integrated deployment remains uncommon on all of them too.
The strongest objection
The best counterargument on Hacker News was temporal rather than denialist. Simon Willison argued that a study whose data mostly ends in 2025 cannot measure the far more capable coding and general agents that users only encountered in late 2025 and early 2026 - the tools that would plausibly displace entry-level work are newer than the window. That is a valid scope objection, and it is also a hypothesis rather than evidence that a break has already occurred. The thread otherwise split between people attributing current losses to pandemic overhiring and people reporting uneven local displacement.
We have covered the claim side of this argument before, where corporate filings told a similar story: the layoff announcements that name AI and the labor data that would show it have not lined up.
The honest caveat
Stanford did not find an AI jobs apocalypse. It found something more uncomfortable: broad labor data still cannot see one, while the first plausible crack is the entry-level ladder - and even Stanford says the post-pandemic economy cannot yet be cleanly separated from AI in that signal. Absence of evidence in aggregate data is genuinely weak evidence of absence when the effect is concentrated in narrow occupations and recent cohorts.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)