AI-Exposed Jobs Were Declining Before ChatGPT — Morgan Frank

Morgan Frank and colleagues tested a claim that has become common, that the job market for AI-exposed work turned down because of ChatGPT. Rather than working from forecasts, they went to the records. Monthly unemployment insurance data from US state offices let them measure unemployment risk by occupation and location, and risk in AI-exposed occupations started rising in early 2022, months before ChatGPT was released. In most occupations and states there is no sharp break at the point ChatGPT arrived.

Two further datasets point the same way. From millions of LinkedIn profiles, graduates from 2021 onward entered AI-exposed jobs at lower rates than earlier cohorts, and that gap opened before late 2022. From millions of university syllabi, graduates who had taken more AI-exposed courses went on to higher first-job pay and shorter job searches after ChatGPT rather than worse outcomes.

The conclusion is about timing rather than about whether AI affects work. Something was already changing in these occupations before generative AI reached the public, and the paper is careful not to claim what. For anyone deciding what to do about AI and their own workforce, it is a useful corrective: the common explanation for the last few years does not fit the dates.