Research update · 2 October 2026
An AI exposure score tells little about whether a firm uses AI
Employment warnings often rank jobs by how much AI could do. A September U.S. Census working paper tested those scores against firms' reported AI use. Exposure was related to adoption, but it explained only about 1–6% of the variation across matched firms. This is a measurement result, not a new count of jobs lost.
What the paper measured
Dey and coauthors combined 14 occupational exposure measures with pre-adoption job mixes from the U.S. Bureau of Labor Statistics. They linked those measures to Census Business Trends and Outlook Survey answers about AI use. The analysis covers matched, single-location U.S. firms with occupation records from 2018–2020 and AI-use responses collected from 2023 into early 2026. Its main revised-question sample contains about 15,500 firm-year observations; the earlier-question sample contains about 57,000.
Exposure predicts some use, not who adopted
Across 14 scores, a one-standard-deviation higher exposure measure was associated with a 4–11 percentage-point higher chance that a firm reported AI use. After controls for year and detailed industry, the range was 1–8 points. Exposure by itself accounted for only about 1–6% of the variation in reported adoption across firms. That last number describes how much variation a score explains across the sample; it is not the probability that any one exposed firm uses AI.
What it cannot answer
The survey asked about AI broadly, including technologies beyond generative AI. Census broadened the question in November 2025 from AI used to produce goods or services to AI used in any business function. The study excludes multi-location firms to align job mix and adoption at the same level, and linking surveys can select which firms remain. It estimates association, not the causal effect of exposure on adoption. It does not measure hires, layoffs, wages or hours.
This result does not erase U.S. early-career hiring and software-vacancy warnings. It narrows the claim they support: a change correlated with task exposure is not automatically a change caused by firms using AI. Employers can also adjust in anticipation of AI without yet reporting use.
The next question for a worker
If a headline calls your role “high risk,” ask what the score measures, whether employers in the study actually used AI, and whether a labor outcome was observed. For your own workplace, ask which tools were introduced and what changed in tasks or staffing. The paper cannot calculate your personal job-loss odds.