Did AI Cause the Hiring Slowdown? What the Job Postings Data Shows

Did AI cause the hiring slowdown in the occupations everyone expected it to hit first? Researchers at the Federal Reserve Bank of New York checked that story against real US job postings and found the timing does not fit. The gap between AI-exposed and AI-safe occupations opened up before ChatGPT was ever released to the public, and did not visibly widen once it arrived. Something has been suppressing hiring in those occupations. The evidence says it started before generative AI existed in anyone's browser.

This matters because the blame story is everywhere and rarely checked. A slowdown gets noticed, a new technology is the obvious headline suspect, and the two get stitched together without anyone lining up the dates. The Fed researchers lined up the dates.

First: a real exposure score, not a guess about a job title

The study combined an occupation-level AI exposure measure built from actual usage patterns with Lightcast's US job-postings data and O*NET task descriptions, then tracked posting volume for high- and low-exposure occupations over time. Fewer than 10 percent of workers and vacancies sit in occupations with an exposure score of 0.4 or higher on that scale, and 40 percent of workers are in jobs with zero measured exposure. Most of the labor market, on this measure, was never in the blast radius to begin with.

Second: the gap predates the technology it gets blamed on

Here is the finding that breaks the simple story. The divergence between high- and low-exposure occupations in posting volume began before 2022, well ahead of ChatGPT's release that November. A cause cannot explain an effect that started before the cause existed. Whatever was already thinning out postings in exposed occupations, it was running before the tool everyone points to was available to a single consumer.

Third: the gap held steady once the technology actually arrived

If generative AI itself were driving the divergence, the natural prediction is a sharper break right around its release, then a widening gap as adoption spread. The researchers found no such break: the gap does not show a clear additional shift in trajectory after 2022, and it stabilizes rather than widens after 2023. Junior roles in exposed occupations did not diverge further from senior ones either, the pattern a straightforward automation story would predict first.

Fourth: what a mistimed cause changes about how you respond

None of this means AI has zero effect on hiring. It means the postings evidence does not support generative AI as the primary driver of the recent slowdown in exposed occupations, which is the conclusion the researchers themselves draw from the timing. That distinction, between a task looking automatable and an actual person losing an actual job, is the same one worth applying to any hiring-slowdown headline you read next. Chasing the wrong cause wastes the one thing worth spending on a real slowdown: building the kind of checkable proof of judgement that gets someone hired regardless of which factor actually thinned the postings. If you want to know how your own role is measured for exposure in the first place, here is how those scores actually get built, and here is a checklist for reading your own week instead of a headline. The check scores six of those judgement domains directly, from your own answers, in about seven minutes.