Research update · 5 October 2026
Job ads changed in ways an occupation risk score can miss
A role can keep its title while its listed tasks change. A U.S. posting study by Wang, Wei and Wang finds both a shift in the mix of advertised roles and a shift in task descriptions within comparable role groups. It does not show that AI caused either change or that a worker lost a job.
What the study measured
The May 2026 preprint samples 9.37 million U.S. Lightcast online postings from January 2021 to June 2025. Its two-model pipeline extracts tasks from ad text and scores how much current generative AI could assist them. The authors group postings by occupation, industry and seniority, then decompose changes in the average exposure score.
From the third quarter of 2023, changes in the mix of posting groups account for 52.01% of the absolute decline in that score. Changes in listed tasks within groups account for 39.46%, and their interaction for 8.54%. Among junior postings the corresponding shares are 60.15%, 18.22% and 21.63%. These are shares of a change in a constructed exposure index, not percentages of jobs lost or employees replaced.
Why it matters
A fixed occupation score may miss changes in the tasks employers advertise within a broad role. That matters when interpreting an entry-level warning: fewer junior ads and different tasks in remaining ads are different outcomes. The study's within-group component compares postings, not the same employer rewriting the same job. Its advertised tasks may not match what workers later do.
What it cannot answer
Average measured exposure rose to an early-2022 peak, before ChatGPT, then fell through 2023 and partly recovered. A post-2023 decomposition cannot isolate AI from the wider hiring cycle or other changes in ad writing. The paper does not link postings to actual employer AI use, completed hires, wages, hours or layoffs. Its model-coded exposure score also depends on the extraction and capability rubric. The authors explicitly do not treat the timing as causal evidence of a product release.
The next useful test would link the same employers' AI adoption, repeated job descriptions and realized worker outcomes, while checking pre-trends and the broader hiring slowdown. Until then, the supported finding is a change in advertised labor demand, not a count of displaced people.