Newshot Introduction
Debates about artificial intelligence’s (AI) impact on the future of work hinge on whether AI will replace workers, augment their productivity, transform the kinds of jobs that already exist, create entirely new jobs, or some combination of these. This Economic Commentary looks at how the rapid deployment of large language models (LLM), beginning with the launch of ChatGPT in late 2022, and followed by subsequent advances in generative AI, are reshaping the US labor market.1 We estimate that since generative AI became widely available, the unemployment rate for workers aged 23–54 has risen by roughly 1 percentage point more in occupations most exposed to AI compared to those in the least exposed, relative to the “pre-AI” baseline period of 2015–2019. This rise in relative unemployment outcomes for occupational groups more exposed to AI’s capabilities is substantive, and the disparity is even larger for workers aged 23–30.
To understand what is driving this shift in labor market dynamics, we look beyond the unemployment rate at labor market flows, that is, how often workers separate from their jobs and how quickly unemployed workers find new ones. Our primary finding is that the rise in relative unemployment is driven primarily by meaningfully lower job-finding rates for workers in highly exposed occupations in the post-AI era compared to rates in the pre-AI baseline period. We do not find consistent evidence of elevated job separation rates among workers in occupations highly exposed to AI. Taken together, these patterns suggest that generative AI may be making it harder, to date, to enter AI-exposed occupations rather than displacing existing workers in these occupations.
Measuring AI Exposure
From recent academic literature, we use three distinct measures of how much a given occupation’s tasks overlap with what AI can do, each reflecting a different approach to this problem. Two of the measures originate from Eloundou et al. (2023), who systematically evaluate more than 19,000 individual tasks catalogued in the US Department of Labor’s O*NET database, a rich compendium of workers’ tasks across 923 occupations. Each task was rated according to a simple criterion, whether an LLM or LLM-powered software system could reduce the time required to complete this task by at least 50 percent while maintaining output quality. This evaluation yields three exposure categories. Tasks with no such potential are classified as “unexposed”; tasks for which a basic LLM interface alone would suffice are classified as “directly exposed”; and tasks requiring additional specialized software built atop an LLM, such as domain-specific retrieval systems or image generation tools, are classified as “indirectly exposed.” From these task-level ratings, the authors construct occupation-level exposure scores. We use their beta score, which credits direct exposure fully and indirect exposure at half weight, reflecting the additional investment required for indirect applications to materialize. Critically, the authors produced two variants of this measure, one based on ratings by human experts familiar with LLM capabilities and one based on ratings by GPT-4 itself. We interpret these scores as measuring the potential for AI to substitute for tasks currently performed by humans. Both are included in our analysis and yield similar results.
A conceptually distinct approach is taken by Handa et al. (2025), who measure the extent to which AI is actually used for tasks associated with a particular job. Analyzing more than four million real-world conversations on the Claude.ai platform, they map observed interactions to O*NET task descriptions and construct occupation-level usage rates. Handa et al. (2025) find that AI use is highly concentrated; software development and writing tasks account for nearly half of all observed interactions, and actual use in very-high-wage occupations (such as for physicians) is notably lower than Eloundou et al.’s (2023) framework would predict, likely because of regulatory barriers and the physical demands of these roles. This use score is a fundamentally different object from what Eloundou et al. (2023) measure. An occupation might be highly exposed in theory yet show low actual AI use because of regulatory constraints, employer inertia, high implementation costs, or slow adoption. Conversely, occupations in which adoption has moved quickly may show high use even if theoretical exposure is moderate. The divergence between these two methodologies can be informative about the extent to which labor market flows are responding to AI’s anticipated effects versus its actual use. We return to this point in the discussion of our analysis results.
How We Track Job Finding and Job Loss across Occupations
To measure how AI exposure affects workers’ odds of finding a job or losing one, we need to follow individuals over time to observe changes in their employment status. This requires repeated observations of the same worker across consecutive months linked to information about the AI exposure of their occupation so that we may compare how these transition rates evolved before and after the widespread availability of generative AI tools across occupations with varying degrees of AI exposure. Our primary data source is the Basic Monthly Current Population Survey (CPS), administered jointly by the US Census Bureau and the Bureau of Labor Statistics (IPUMS, Flood et al., 2024).2 The CPS interviews households for four consecutive months, rotates them out for eight, and then interviews them for four more. This panel structure allows us to link individuals across adjacent months and directly observe transitions between employment states.
We restrict the sample to civilians aged 23–54, excluding most younger workers still in school and older workers near retirement whose labor market behavior may reflect lifecycle considerations unrelated to AI. We focus on the unemployment rate along with two transition outcomes. An unemployed worker is classified as a “job finder” if they are employed in the subsequent month. For example, if a person is unemployed in January and employed in February, we say that they are a job finder in February. Similarly, an employed worker is classified as “separated from their job” if they are unemployed or out of the labor force in the subsequent month. These individual flows, grouped at the occupational level (according to their AI exposure), form the dependent variables in our analysis.
Because unemployed workers do not currently hold jobs, they cannot be assigned to an occupation in the conventional sense. The CPS surveys all unemployed respondents about the occupation of their most recent job, and we use this information to classify unemployed workers into occupational groups by AI exposure level.3,4
Empirical Strategy
We group the 923 O*NET occupations into 20 equal-sized groups (ventiles) ranked by AI exposure according to each of our three measures. We then estimate, for each ventile group, the change in the average unemployment rate, job-finding rate, and job-separation rate between a pre-AI baseline period and a post-AI comparison period.
Our pre-AI baseline period is January 2015 through December 2019. This window predates the public availability of generative AI tools and is also free of COVID-19-pandemic-related distortions to the labor market, distortions which were severe and idiosyncratic during 2020–2022. Our comparison period for unemployment covers May 2025 through April 2026, a period in which LLMs had become widely available and extensively adopted by businesses and individuals (Bick, Blandin, and Deming, 2026). For job-finding and job-separation rates, we extend our comparison period to include January 2024 through April 2026.5 Because of its rotating panel structure and attrition in the CPS, unemployed individuals with the requisite month-to-month linkages are a smaller subset of the sample used to compute unemployment rates. Extending the comparison period allows us to regain some of this lost sample size and maintain the statistical precision of the patterns documented below.6
To see how AI has affected the labor market, we plot changes in the unemployment rate and the associated flow rates for each occupational group against its AI exposure ranking. We compare how much the labor market has changed before versus after the advent of generative AI across occupations with varying degrees of exposure.7 The key question is whether unemployment or flow rates changed more in occupations in which AI can do more of the work. If AI is reshaping the labor market, we would expect the biggest changes to show up precisely in those occupations.
AI Exposure and Labor Market Outcomes
Figure 1 presents our headline result: the relationship between AI exposure and changes in the unemployment rate across our 20 occupational groups, post- minus pre-AI availability, for each of the three exposure measures. All three measures show a positive relationship. Occupations with higher AI exposure saw larger relative increases in unemployment than occupations with lower exposure. The pattern is clearest for Eloundou et al.’s (2023) LLM-rated task exposure metric (panel (b)), where the fitted line climbs from about -0.6 percentage points at the lowest exposure occupations to about 0.7 percentage points at the highest, giving a 1.3 percentage point predicted gap between least- and most-exposed occupations (see bold number to the right of the panel). Their human-rated task exposure metric (panel (a)) shows a similar positive slope with more scatter around the fitted line. Handa et al.’s (2025) usage-based measure (panel (c)) also shows a positive slope, though slightly flatter compared to the slopes of the other measures.