Research check · 14 October 2026
Higher sector employment does not clear automatable jobs
A U.S. working paper finds about 4% higher relative employment by 2024 in state-industry groups more exposed to tasks that AI could help workers perform. It finds no statistically detectable employment difference for exposure to tasks AI could do more independently. The comparison is based on task exposure, not observed AI adoption. It cannot say that AI created those jobs or that no worker lost one.
What the study measured
Johnston and Makridis compare U.S. state-by-industry groups from 2017 to 2024. They use occupations' pre-2020 task mix to assign exposure to large language models, then compare employment and output trends across groups. The employment data come from the Quarterly Census of Employment and Wages. One standard deviation more exposure to augmentable tasks corresponds to roughly 4% higher relative employment by 2024. The automatable-task employment estimate is not statistically distinguishable from zero.
Where the claim stops
Exposure is a prediction about tasks, not a record that an employer installed or used AI. The authors control for state-year, industry-year and remote-work differences, but unmeasured local changes could still track exposure. Their automatable-task employment series had a positive trend before the assumed 2021 rollout; they adjust for it. That makes “AI did not cut automatable work” too strong. The data end in 2024 and do not count layoffs, completed hires or later worker moves.
Does this change the answer?
No. The study adds a sector-level counterpoint, not a clean reversal of the narrower hiring evidence. Copilot hiring studies use different firm samples and outcomes; a sector employment gain could coexist with weaker entry routes inside a sector. The 6.1 million vulnerability estimate remains a conditional map, not a job-loss count. A stronger test would observe AI use and follow hires, separations and pay in the same employers.