AI Labor Tracker

AI-Exposed Workers Aren't Losing Jobs Faster. They're Taking Weeks Longer to Find the Next One.

A Cleveland Fed study published today tracked workers month to month. In the occupations most exposed to AI, people aren't separating from jobs any faster than before. They're finding new ones more slowly, and the young are hit hardest.

This morning the Federal Reserve Bank of Cleveland published a short paper with a plain title, “Job Finding and Unemployment in the Age of AI.” Its authors, Alexander Cline and Barış Kaymak, set out to answer the question most AI-and-jobs arguments skip.

Several studies have found that unemployment has risen more in jobs AI can do. Cline and Kaymak estimate the gap at about one percentage point for workers aged 23 to 54, compared with before generative AI arrived. What nobody had pinned down was how it happened.

There are only two ways to raise unemployment in a group of jobs. More people can lose those jobs, or the people who lose them can take longer to get hired again. Those are very different problems, with very different fixes.

The Cleveland Fed’s answer is the second one. “Worker movements in and out of unemployment reveal this is not because workers in at-risk occupations are separating from their jobs at a faster pace,” the paper says, “but because they are taking longer to find new ones.”

So the question I want to answer is simple. If AI is hurting workers, where exactly is the damage happening, and what should someone in an exposed job do about it?

How they measured it

The study uses the Current Population Survey, the monthly government survey of about 60,000 households that produces the unemployment rate. Households stay in the survey for four months in a row, so you can see the same person go from employed to unemployed, or from unemployed to employed.

Cline and Kaymak sorted 923 occupations into 20 equal groups by how exposed their tasks are to AI. They used three different exposure measures. Two come from a 2023 study by OpenAI and University of Pennsylvania researchers that rated each task on whether a language model could cut the time it takes by at least half without hurting quality. One version was rated by people, one by GPT-4. The third comes from Anthropic’s analysis of what people actually use its Claude chatbot for.

Then they compared each group’s unemployment and job-flow rates in a “pre-AI” period, 2015 through 2019, with the recent period. For unemployment that’s May 2025 through April 2026. For the job flows, which need people tracked across two months, it’s January 2024 through April 2026.

One detail I appreciate: the national unemployment rate for 23- to 54-year-olds was about 4.5% in both periods, so the comparison isn’t just picking up a weaker economy.

Where the damage is

On the way out of jobs, they found nothing consistent. One exposure measure showed no relationship at all between AI exposure and the rate at which people left jobs. The other two showed slight relationships pointing in opposite directions. “None of these patterns are statistically robust,” the authors wrote.

On the way back in, the pattern was clear. In the occupations most exposed to AI, by the two task-based measures, the monthly job-finding rate for unemployed workers fell about six percentage points more than in the least exposed.

That sounds small until you know the baseline. In a typical month about a third of unemployed workers find a job.

By the paper’s own arithmetic, that drop stretches the implied average spell of unemployment from about 10.5 weeks to more than 13.

That’s two or three extra weeks without a paycheck for the average person in those jobs. Averages hide a lot, so some people will wait much longer.

The usage-based measure, built from what people actually do with Claude, showed the same direction but a flatter slope. The authors offer a reason that makes sense to me. Actual use mixes two things: AI doing the task instead of a person, and AI helping a person do the task faster. Citing Anthropic’s research, they note that roughly 43% of observed use looks like substitution and 57% like augmentation. In jobs where AI mostly helps, a worker can become more valuable to hire.

The young get the worst of it

The age split is where this gets personal. For workers aged 23 to 30, moving from the least to the most exposed occupations raises the predicted unemployment rate by about two percentage points. For workers 45 to 54, the gap is at most one point.

The authors’ explanation is the one I’d give too. Younger workers compete for entry-level positions, “which are specifically the roles in which AI tools may be most quickly substituting for routine cognitive tasks.” Older workers carry expertise, networks and knowledge of a specific company that are harder to replace.

There’s a detail in the footnotes worth knowing. The young-worker results leave out a person’s very first spell of unemployment when entering the labor force, because they don’t have a previous occupation to classify. So this isn’t only about new graduates who can’t land a first job. It’s young people who already had a job in an exposed occupation, lost it, and are struggling to get the next one.

That fits with research from another Fed bank. In August the Richmond Fed’s Katarína Borovičková and Claudia Macaluso published a brief looking at who accounts for the national slowdown in job finding. They found brand-new labor force entrants were not “the margin where the action is.”

The biggest drop, they found, was among workers with the most stable job histories, the ones who are almost always employed. From November 2022 to September 2025, the job-finding rate for that group fell 13 percentage points, against 2 points for workers who cycle in and out of jobs more often. “We have not seen such a significant difference in any of the previous recessions, not to mention when the economy is expanding,” they wrote. And within the unemployed, those in highly AI-exposed occupations saw the largest declines.

Three studies, one door

This isn’t the first time this site has written about the re-entry problem. On November 2 we covered a Goldman Sachs analysis of workers laid off from AI-exposed jobs, which found they took about a month longer to find new work than other displaced workers.

What’s new today is the other half of the picture. Goldman looked at people already displaced. The Cleveland Fed looked at everyone in these occupations and asked whether they were being displaced more often. The answer was no, not yet.

Put the three together and they point the same way, using different data and different methods. The Cleveland and Richmond studies use monthly survey flows; Goldman used the government’s displaced worker supplement. All three find the damage concentrated at the moment of getting hired.

It also squares with the government’s job turnover data. In September, employers’ layoff and discharge rate was 1.0% of employment, according to the Bureau of Labor Statistics’ JOLTS report, inside the narrow range it has held since mid-2024, as Indeed Hiring Lab noted. Firing is not where the action is.

Why the door, not the exit

I think the simple explanation is the right one here. For an employer, not filling a job is far cheaper and easier than eliminating one.

A layoff comes with severance, legal review, sometimes a WARN notice, a hit to morale, and a news story. Letting a position go unfilled after someone quits comes with none of that. If a manager thinks AI might handle part of a role in a year or two, the low-risk move is to wait. The Cleveland authors say the same thing in economist language: firms appear to be “using attrition and reduced hiring as a low-risk way to shrink their exposure to labor costs.”

That’s also why the headlines and the data keep talking past each other. A layoff is an event with a date, a company name and a number. A job that was never posted isn’t an event at all. Nobody announces it, and no survey of layoffs can count it.

The people who pay for it are the ones on the outside at the wrong time. “For many workers already inside these occupations, this may feel like business as usual,” the paper says. For people trying to get in, especially younger workers, “the door appears to be harder to open.”

What this doesn’t prove

To be honest with you, this paper doesn’t prove AI caused any of it, and the authors don’t claim it does.

Exposure measures describe what AI could do in an occupation. They don’t show what employers actually did. Occupations that score high on AI exposure, like programming and financial analysis, may also have been hit by other things since 2022, such as higher interest rates squeezing tech and finance hiring. The authors’ method compares exposed and unexposed occupations against their own pre-2020 baselines, which helps, but it can’t rule out every story.

Two other limits matter. Unemployed people are classified by the occupation of their last job, and many of them search in other fields. The authors think that likely makes their estimates too small. And the comparison period ends in April 2026, so it’s already six months old.

The paper also carries the usual label: “This paper and its data are subject to revision.”

My take

My read: the AI labor shock so far is a hiring problem, not a firing problem. It is real, it is measurable, and it lands hardest on younger workers trying to move from one job to the next.

That matters for what we do about it. Severance rules, layoff notices and retraining money for the laid-off are all aimed at the exit. None of them help someone who was never fired but can’t get hired. If the damage is at the door, the fixes have to be at the door too: apprenticeships, paid trial periods, internal moves before outside hires, and employers being honest with themselves about whether a role is paused or gone.

I care about this because of what a job gives a person. On my podcast in 2022 I said, “There’s an incredible amount of dignity that you get when people can take care of themselves and their families.” A few extra weeks of searching sounds minor in a research paper. It’s not minor when it’s your rent.

What to do if you’re in an exposed job

The Cleveland paper ends on uncertainty, and I think that’s right. The authors note that general-purpose technologies have historically “raised productivity and labor demand broadly, even as they eliminated specific types of work in the short run,” and that whether AI follows that pattern “remains an open question.”

Until we know, the practical lesson from this research is narrow and useful. In the jobs most exposed to AI, the risk isn’t concentrated in keeping your job. It’s concentrated in changing it.

My recommendation: if you work in one of those occupations and you’re thinking about leaving, line up the next job before you give notice, and start the search months earlier than you think you need to. The data says the gap between jobs is where people in your position are losing ground, so keep that gap as short as you can.