The Steadiest Workers Lost 13 Points of Job-Finding. The Churn-Prone Lost 2.
AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)
On 12 August 2026, the Federal Reserve Bank of Richmond published an Economic Brief on who is absorbing the decline in job-finding rates. The answer is not the group you would expect from a normal downturn, and it is not the group most AI-and-jobs coverage has been looking at.
From the November 2022 peak to the September 2025 trough, the job-finding rate fell 13 percentage points for workers the brief classifies as the primary type — people with stable, long-tenure employment histories who are almost always working. Over the same period it fell 2 points for the secondary type, the group that historically churns in and out of work and normally absorbs most of the cyclical damage.
That is close to backwards from how a labour market usually behaves.
The number that is not about AI, and the finding that is
This needs stating carefully, because it is the easiest thing in this brief to get wrong.
The 13-points-versus-2 comparison is about labour market attachment. It sorts workers by how stable their employment histories are, using a latent-type framework from a 2023 working paper by Hie Joo Ahn, Bart Hobijn and Ayşegül Şahin. It is not a comparison between AI-exposed and non-exposed occupations, and anyone who tells you the Fed found that AI-exposed workers lost 13 points of job-finding while everyone else lost 2 has merged two different charts.
The AI-exposure finding sits in the same brief and is reported differently. Sorting workers into quartiles by how much their occupation's tasks overlap with capabilities described in AI patents, the authors find that outflow rates used to move together across all four quartiles, tracking the business cycle with only small level differences. Since 2023, they diverge — and in the brief's own words, "workers in highly AI-exposed occupations have seen the largest declines in the job-finding rate."
No percentage-point figure is published for that cut. The finding is directional, and we are reporting it as directional.
Where the two dimensions meet is the brief's own conclusion: "Most of the recent decline in unemployment outflow rates is being borne by strongly attached workers, particularly those whose occupations are exposed to AI."
Why this is not a layoff story
A companion brief from the same two authors, published a week earlier on 5 August 2026, establishes the frame. The recent rise in long-term unemployment is almost entirely accounted for by the decline in the rate at which unemployed workers leave the unemployment pool. Inflows into unemployment from employment remain low by historical standards. There has been, in that brief's words, no layoff spike comparable to those observed in 2008 to 2009 or in 2020.
So the mechanism under discussion is not a firing wave. It is a hiring problem. People are not being pushed out of work in unusual numbers; the ones who are out are taking much longer to get back in, and the people taking longest are the ones whose CVs suggest they should be taking the least time.
That distinction matters for anyone deciding what to do about it. Job security and job replaceability are different risks, and this data speaks to the second one. Nothing here says your current job is at elevated risk. It says that if you lose it, the re-entry may be slower than your own employment history would predict.
The historical comparison, and its limits
Job-finding rates for the primary type always fall furthest in a downturn. That part is ordinary. What the brief flags is the size of the gap.
| Episode | Primary type | Secondary type |
|---|---|---|
| Great Recession, peak to trough | -19 points | -10 points |
| November 2022 to September 2025 | -13 points | -2 points |
Exhibit — peak-to-trough fall in job-finding rates, by worker type
The brief states: "We have not seen such a significant difference in any of the previous recessions, not to mention when the economy is expanding."
Two caveats belong next to that table. These are peak-to-trough comparisons across business cycles of different lengths and different causes, and the brief reports no statistical significance test or confidence interval for the comparison. The underlying data is the Census Bureau's Current Population Survey, which is a household survey subject to sampling error, and no margin of error is published for the group-level splits shown.
What the classifications actually measure
Both of this brief's key dimensions are proxies, and both are worth understanding before quoting the results.
AI exposure follows Michael Webb's 2019 method: score an occupation by how much of its task content overlaps with the technical capabilities described in AI patents. That measures what AI has been designed to do, not what AI has been deployed to do, and not whether anyone in that occupation has actually been displaced. An occupation can score high on patent overlap and have almost no real AI in its workplaces, or the reverse.
Worker type is inferred, not observed. Nobody is asked whether they are primary or secondary; the model assigns them from employment histories. We did not independently open either underlying paper, and we are relying on the brief's own description of both methods.
One more thing the brief rules out on its own evidence: new entrants to the labour market are not where this is happening. Their job-finding rates have declined too, but the drop is modest relative to job losers and job leavers, and the authors set entrants aside for the rest of the analysis. The popular version of the AI-and-jobs story — that graduates are being locked out first — does not get support here.
What would change this picture
The authors have already named the missing piece. This is the second article in a three-part series, and the third is set to address the question this one leaves open: how much of the aggregate decline in job-finding rates is common to all worker groups, and how much is specific to AI-exposed occupations. Until that decomposition exists, the honest position is that AI exposure predicts where the decline is worst without anyone having shown it causes it.
Two other things would move this. A reversal in the exposure-quartile divergence would suggest 2023 to 2025 was a composition effect rather than a technology effect. And a rise in inflows to unemployment would change the story entirely — that would be the layoff wave this data currently does not show.
Source: Katarína Borovičková and Claudia Macaluso, "Worker Types, AI Exposure and the Recent Decline in Job-Finding Rates," Federal Reserve Bank of Richmond Economic Brief, August 2026, No. 26-26. The companion brief is "The Ins and Outs of Long-Term Unemployment," No. 26-25, August 2026. Views expressed in those articles are those of the authors and not necessarily those of the Federal Reserve Bank of Richmond or the Federal Reserve System.
Frequently asked questions
What is a job-finding rate, and why is it different from a layoff number?
A job-finding rate measures how likely an unemployed person is to move into a job in a given period. It describes the exit from unemployment, not the entry into it. A layoff statistic counts people losing jobs; a job-finding rate describes how hard it is to get the next one. The two can move in opposite directions, and right now they are: the Richmond Fed reports that inflows into unemployment remain low by historical standards while job-finding rates have fallen sharply. That combination means rising unemployment driven by people staying unemployed longer, not by more people being let go.
Did the Richmond Fed say AI caused the decline in job-finding rates?
No. The brief reports that since 2023, workers in occupations highly exposed to AI have seen the largest declines in job-finding rates, and describes that pattern as "consistent with" a body of research on automation and displacement. It does not claim causal identification, and the authors note that pre-2023 differences across exposure groups likely reflect correlations between AI exposure and other occupational characteristics rather than AI itself. A third article in the series is planned to separate how much of the decline is common to all workers versus specific to AI-exposed occupations.
What are primary and secondary worker types?
They are statistical classifications, not job titles. The brief uses a latent-type framework from a 2023 working paper by Hie Joo Ahn, Bart Hobijn and Ayşegül Şahin, which sorts workers into three groups by how attached they are to the labour market: a primary type (about 55 percent of the population, almost always employed), a secondary type (about 14 percent, strongly attached but unemployed more often), and a tertiary type (about 31 percent, weakly attached and often out of the labour force). These are model-derived groupings inferred from employment histories, not labels anyone carries or self-reports.
Which occupations count as highly exposed to AI?
The brief follows a 2019 working paper by Michael Webb, which scores occupations by the share of their tasks that overlap with the technical capabilities described in AI patents, then sorts them into quartiles. The brief names computer programmers, financial analysts and engineers as highly exposed, and construction workers, food service workers and personal care aides as less exposed. This is a measure of task overlap with what AI patents describe, not a measure of whether AI has actually been deployed in a particular workplace or has actually replaced anyone.
Is it normal for stable workers to be hit hardest?
Job-finding rates always fall further for the primary type in a downturn, so the direction is not new. The size of the gap is. In the Great Recession, job-finding probability fell 19 percentage points for the primary type and 10 points for the secondary type from peak to trough. In the current episode, from the November 2022 peak to the September 2025 trough, it fell 13 points for the primary type and 2 for the secondary. The brief states that it has not seen such a significant difference in any previous recession, let alone during an expansion.