AI and the Composition of Labor Demand: Evidence from Online Job Postings
Published in Working Paper, 2026
Working paper: not yet peer reviewed.
This paper examines where the post-2022 contraction in labor demand fell in relation to occupational AI exposure, by analyzing online job postings across the European Union from 2021 to 2025. Since the release of ChatGPT, postings for the most AI-exposed occupations have declined by 19–29% relative to the least exposed ones, or 4–7% per standard deviation of exposure, with the gap widening each quarter. The pattern holds across five measures of occupational AI exposure, survives controls for interest rate sensitivity, the pre-ChatGPT hiring run-up and teleworkability, and is partially replicated in an out-of-sample test on Australian data.
Within occupations, job skills that are most similar to AI capabilities are mentioned 12% less frequently in job descriptions, suggesting changes in task composition. Industry-level analysis shows no corresponding change in labor or capital productivity, consistent with “so-so automation”. Finally, the association between AI exposure and posting growth follows an inverted U-shape in experience: it is most negative for entry-level postings and for those requiring over ten years of experience.
An interactive dashboard accompanies the paper with headline estimates, event studies, the horse race against alternative explanations, and an occupation explorer.
Recommended citation: Demirev, G. (2026). "AI and the Composition of Labor Demand: Evidence from Online Job Postings." Working paper, SSRN 7054839.