Joint with Bharat Chandar
New technologies can raise demand for some workers while lowering it for others, but knowing which workers gain and which lose is challenging. We study this question for generative AI. Our model decomposes changes in labor demand following firm AI adoption into a firm-wide productivity effect and worker-type-specific changes in task content. We take the model to the data, analyzing a sample of 1.25 billion job postings and 154 million employment records across 41 countries. We infer AI adoption from job advertisements that involve generative AI use. An instrumented event study shows that foreign affiliates of AI-adopting companies reduce the junior share of their workforce relative to comparable control affiliates. The decline primarily comes from growth in senior employment rather than from a fall in junior employment, with suggestive evidence of modest overall employment growth. Senior employment shifts toward AI-exposed occupations, while our point estimates suggest a shift away from these occupations among juniors. The model-based decomposition provides suggestive evidence of modest productivity gains from AI adoption.
PDF | SSRN | Media: Bloomberg | Media: Forbes | Media: FAZ
Joint with Kartik Akileswaran, Andrei Andronic, and Aristotle Vossos
Educational investments determine the future supply of skills. We examine how generative AI is changing students' decisions about what to study. Linking millions of LinkedIn profiles to UK application records, we measure the AI exposure of university-subject pairs based on the occupations of their past graduates, and follow applications to 1,090 degrees over the 2020-2025 cycles. Human capital theory predicts that students should leave subjects whose prospects are eroded by AI, but we find the opposite. After the release of ChatGPT, applications to a degree one standard deviation higher in exposure grew six percent more than applications to less exposed degrees. The shift appears among men and women and across ethnic groups, while total demand for university places is unchanged. Our labor market analysis shows that the job opportunities for juniors have deteriorated for more exposed degrees. Students are thus moving toward the subjects in which entry-level job opportunities are declining.
PDF | SSRN | Media: Financial Times | Media: King's College London | Blog
Joint with Philip Tomei
Which jobs can AI learn to do? We examine this for every occupation in the US economy. Existing indices measure the overlap between AI capabilities and occupational tasks rather than which tasks AI systems can learn to perform, and as a result misclassify occupations where the gap between present capability and learnability is large. Reinforcement learning in post-training, now the dominant paradigm at the frontier, is structured around task completion and maps more directly onto the task-based architecture of occupational classifications than prior approaches. Using LLM annotators guided by a rubric developed with RL experts and validated against confirmed deployment cases, we score all 17,951 ONET tasks for training feasibility and aggregate to the occupation level, producing an RL Feasibility Index. The index diverges sharply from existing AI exposure measures for specific occupation groups: power plant operators, railroad conductors, and aircraft cargo handling supervisors score high on RL feasibility but low on general AI exposure, while creative and interpersonal roles (musicians, physicians, natural sciences managers) show the reverse. These divergences carry direct implications for policy interventions.
PDF | SSRN | Media: American Enterprise Institute
Joint with Daniel Carey
Occupations with identical AI exposure can experience opposite labor market trajectories depending on which tasks are automated. Building on Autor and Thompson (2025), we formalize this insight in a model that separates two margins of automation: displacement on the labor-demand side, and a supply-side expertise-threshold channel that expands or contracts the pool of qualified workers. The model predicts that these margins interact, with rising expertise requirements dampening employment declines from displacement. We test these predictions using hundreds of millions of job postings across 39 countries and exploiting ChatGPT's November 2022 release as a natural experiment. We find that expertise-raising automation leads to a small but significant increase in advertised wages, while AI exposure reduces job postings by 6.2 percent, with suggestive evidence that effects are larger in countries with stricter employment protection and lower digital readiness. The employment interaction is positive and significant, confirming that the employment effect of expertise changes depends on the degree of automation displacement.
PDF | SSRN | Media: Financial Times | Media: The Times | Media: Windfall Trust
Bouke Klein Teeselink
This paper examines the effects of LLMs on UK labor market outcomes. We use a difference-in-differences design that compares outcomes across firms and occupations from 2021 to 2025 based on their differential exposure to LLM capabilities. Our results show that highly exposed firms reduce employment, with effects concentrated in junior positions. These firms sharply curtail new hiring, with technical and creative roles experiencing the steepest declines. At the occupation level, roles more exposed to LLMs also show substantial reductions in job listings. The observed effects concentrate almost entirely in high-wage segments. Hence, our findings suggest that LLMs may compress wage inequality.
PDF | SSRN | Media: Guardian | Media: Reuters | Media: Forbes | Media: Financial Times | Media: The Economist | Media: The Independent | Media: FAZ | Media: King's College London
Joint with Matthew R. Jordan, Nicholas Adolph & Shane Frederick
Patients are often uncertain about the cost of a medication until they arrive at the pharmacy to pick it up. The ultimate out-of-pocket cost for patients is sometimes reduced by copay cards: manufacturer-issued coupons sent to pharmacies, physicians, or patients. Using a data set containing transactions from ~85% of US pharmacies, we find that copay cards increase pick-up rates of prescribed medications even when the ultimate out-of-pocket price remains the same. In other words, copay card discounts not only increase consumption by lowering price, they shift the entire demand curve by introducing transaction utility to the purchase decision, thereby increasing demand at every price level.
Joint with Georgios Melios
This paper provides causal evidence on the impact of religious identification on political preferences, gender norms, societal beliefs, and group behavior. Exploiting clergy sexual abuse scandals as a source of exogenous variation in Catholicism, we demonstrate that religious de-identification leads to significant shifts in individual attitudes and political alignment. Using data on millions of U.S. college freshmen and county-level voting records, we find that secularization causes more progressive positions on issues like abortion rights and same-sex marriage, but more conservative views on universal healthcare and military spending. The net effect is a substantial leftward shift in overall political orientation. We also document more progressive gender attitudes, particularly regarding women's workforce participation, among those who deidentify as Catholic. Notably, religious de-identification reduces individuals' propensity to engage in other group activities. Additional analyses suggest that as individuals disaffiliate from Catholicism, they increasingly identify with their social class, which polarizes economic preferences between high and low-income groups.
Joint with Georgios Melios
Disability benefits typically reduce labor force participation through raising non-labor income (income effects) and phasing out benefits with earnings (substitution effects). Distinguishing these mechanisms matters for policy design, as income effects are non-distortionary while substitution effects generate deadweight loss. We isolate pure income effects for the UK's Personal Independence Payment (PIP) program, which supports disability-related costs without work restrictions or earnings tests. Following a 2013 reform, we use differential eligibility changes across health conditions and quasi-random regional assignment to different assessment providers in a difference-in-differences design with matched panel data. Despite substantial benefit receipt changes (1.8-5.9 percentage points across strategies), we only find positive employment effects in stricter assessment regions. For individuals who gain or lose benefits due to condition-specific eligibility changes, the effects are close to zero. These findings suggest that disability benefits designed as cost-of-living supplements can provide financial security without necessarily disincentivizing work.
PDF | SSRN | Media: BENEFITS
Work in Progress
Joint with Scott Dickinson