人工智能暴露工人的可再培训性如何？（英）

N O .1 1 6 5A U G U S T2 0 2 5 Benjamin Hyman|Benjamin Lahey|Karen Ni|Laura Pilossoph How Retrainable Are AI-Exposed Workers?Benjamin Hyman, Benjamin Lahey, Karen Ni, and Laura PilossophFederal Reserve Bank of New York Staff Reports, no.1165August2025 Abstract We document the extent to which workers in AI-exposed occupations cansuccessfully retrain for AI-intensive work. We assemble a new workforcedevelopment dataset spanning over 1.6 million job trainingparticipationspells from all U.S.Workforce Investment and Opportunity Act programsfrom 2012-2023linked with occupational measures of AI exposure.Using earnings records observed before and aftertraining, we comparehigh AI exposure trainees to a matched sample of similar workers whoonlyreceived job search assistance. Wefind that the average earningsreturn to training among AI-exposedworkers is high, around$1,470 perquarter. Low-exposure trainees capture higher returns, and traineeswhotarget AI-intensive work face a 29percentearnings return penalty relative totheir high exposurepeers who pursue more general training. We estimatethat between 25to 40percentof occupations are JEL classification:J08, M53, O31Keywords:artificial intelligence, active labor market policies, job training, labor markets Hyman: Federal Reserve Bank of New York (email:ben.hyman@ny.frb.org).Lahey:New YorkUniversity, Department of Economics(email:bpl9631@nyu.edu).Ni: Harvard Kennedy School(email:kni@g.harvard.edu). Pilossoph:Duke University (email: laura.pilossoph@duke.edu).The authorsthankJoe Altonji, David Card, Lisa Kahn, Fabian Lange, Michael Lee,Kyle Myers,Steve Raphael, DanielRock, Wilbert van der Klauuw, Till von Wachter, and seminar participants at the annual SOLEconference,Federal Reserve Banks of San Francisco and New York, and UC Berkeley IRLE for helpfulcomments onan earlier version of the paper.Theyalso thank Jonathan Lee for excellent researchassistance. Ni gratefullyacknowledges support from the Institution of Education Sciences, U.S. This paper presents preliminary findings and is being distributed to economists and other interestedreaders solely to stimulate discussion and elicit comments. The views expressed in this paper are those ofthe author(s) and do not necessarily reflect theposition of the Federal Reserve Bank of New York or the To view the authors’ disclosure statements, visithttps://www.newyorkfed.org/research/staff_reports/sr1165.html. 1Introduction The debate over whether advances in artificial intelligence (AI) will ultimately complementor substitute labor has drawn significant attention (Autor (2024); Deming et al. (2025); Hampole et al. (2025)). However, there is a dearth of research examining the role that existingjob training programs might play in helping AI-exposed workers adapt to the evolving labor Thisshort paper evaluates the effectiveness of the US’s flagship workforcedevelopment program—the Workforce Innovation and Opportunity Act (WIOA), formerlythe Workforce Investment Act (WIA)—in helping workers transition out of jobs facingAI-related pressure and into jobs with higher AI complementarity.We assemble a noveldataset of over 1.6 million individual WIOA/WIA training spells from 2012 to 2023, linkedto administrative earnings records spanning several quarters before and after training. The We find that earnings returns to training for AI-exposed workers are large andpositive,but workers capture higher returns when they avoid targeting AI-intensiveoccupations in their next jobs. In an ideal experiment, these returns would be identified by random assignment. Absent random assignment, we follow the approach of Rothstein et al.(2022), matching each WIOA/WIA trainee to a control worker who only received jobsearch assistance. The credibility of this design is strengthened by the fact that, although outcomes for workers displaced from high–AI exposure occupations, and for those who aretargetingtransitions into AI-intensive occupations after training.Thatis,we canseparately estimate the effects of more general (non-AI) and specific (AI-deepening) skill We have four main findings. (i) The modal participant displaces fromtop quintileAIexposure occupations despite being relatively low income.This is not driven by thecomposition of the sample—training participants have composition similar to the CPSUnemployed and Displaced Worker samples.(ii) The mean earnings returns for trainingparticipants displaced from above-median (“high”) AI exposure occupations are large andpositive—around$1,470 per quarter relative to the control group and only about 25%lower than the returns for low AI exposure workers.(iii) Participants who target high AI Using the matched sample, we also construct an AI Retrainability Index (AIR) whichranks occupations by the share of workers who can successfully retrain into higher-wage, moreAI-intensive roles after leaving those occupations. Defining AI retrainability in terms of bothskill and earnings helps mitigate survivorship b