My research focuses on three topics. First, I study the distributional implications of new production technologies using micro data that link workers, firms, and robots. Second, I explore the technology adoptions that drove the Second Industrial Revolution using newly-digitized microdata from the Census of Manufactures. Finally, I evaluate whether educational policies can help workers adjust to new technologies.
Working Papers
The Adoption of ChatGPT, with Emilie Vestergaard.
Coverage: Becker Friedman Institute,
Abstract: We study the adoption of ChatGPT, the icon of Generative AI, using a large-scale survey experiment linked to comprehensive register data in Denmark. Surveying 100,000 workers from 11 exposed occupations, we document ChatGPT is pervasive: half of workers have used it, with younger, less experienced, higher-achieving, and especially male workers leading the curve. Why have some workers adopted ChatGPT, and others not? Workers see a substantial productivity potential in ChatGPT, understand it substitutes for human expertise, and expect little cross-task substitution. Adoption is hindered by practical hurdles, including employer restrictions and required training, rather than existential fears of job redundancy or technology dependency. Informing workers about expert assessments of ChatGPT shifts workers' beliefs and intentions but has limited impacts on their adoption of the technology.
Gaining Steam: Incumbent Lock-in and Entrant Leapfrogging, with Richard Hornbeck, Shanon Hsuan-Ming Hsu, and Martin Rotemberg. Submitted.
Abstract: We examine the long transition from water to steam power in US manufacturing, focusing on early users of mechanical power: lumber and flour mills. Digitizing Census of Manufactures manuscripts for 1850-1880, we show that as steam costs declined, manufacturing activity grew faster in counties with less waterpower potential. This growth was driven by steam powered entrants and agglomeration, as water powered incumbents faced switching barriers primarily from sunk costs. Estimating a dynamic model of firm entry and steam adoption, we find that the interaction of switching barriers and high fixed costs creates a quantitatively important and socially inefficient drag on technology adoption. Despite substantial entry and exit, switching barriers remained influential for aggregate steam adoption throughout the 19th century, as water power required lower fixed costs and therefore was attractive to relatively low productivity entrants. These entrants then became incumbents, locked into water power even if their productivity grew.
Changing Tracks: Human Capital Investment after Loss of Ability, with Pernille Plato and Jakob R. Munch.
Revision Requested, American Economic Review.
Coverage: Becker Friedman Institute, Chicago Booth Review [video], Forked Lightning, University of Copenhagen.
Abstract: We provide the first evidence on how workers invest in human capital after losing ability. Using quasi-random work accidents in Danish administrative data, we find that workers enroll in bachelor's programs after physical injuries. Exploiting differences in eligibility driven by prior vocational training, we find that higher education moves injured workers from disability benefits to full-time employment. Reskilled workers earn 25% more than before their injuries and avoid ending up on antidepressants. Reskilling subsidies for injured workers pay for themselves four times over, and current rates of reskilling are substantially below the social optimum, especially for middle-aged workers.
What Works for the Unemployed? Evidence From Quasi-Random Caseworker Assignments, with Jakob R. Munch and Mette Rasmussen.
Revision Requested, Journal of Political Economy.
Abstract: This paper examines if active labor market programs help unemployed job seekers find jobs using a novel random caseworker instrumental variable (IV) design. Leveraging administrative data from Denmark, our identification strategy exploits that (i) job seekers are quasi-randomly assigned to caseworkers, and (ii) caseworkers differ in their tendencies to assign similar job seekers to different programs. Using our IV strategy, we find assignment to classroom training increases employment rates by 25% two years after initial job loss. This finding contrasts with the conclusion reached by ordinary least squares (OLS), which suffers from a negative bias due to selection on unobservables. The employment effects are driven by job seekers who complete the programs (post-program effects) rather than job seekers who exit unemployment upon assignment (threat effects), and the programs help job seekers change occupations. We show that job seekers exposed to offshoring -- who tend to experience larger and more persistent employment losses -- also have higher employment gains from classroom training. By estimating marginal treatment effects, we conclude that total employment may be increased by targeting training toward job seekers exposed to offshoring.
Robot Adoption and Labor Market Dynamics
Abstract: I study the distributional impact of industrial robots using administrative data that link workers, firms, and robots in Denmark. I estimate a dynamic model of how firms select into and reorganize production around robot adoption. I find that firms expand output, lay off production workers, and hire tech workers when they adopt robots. I embed the firm model into a general equilibrium framework that endogenizes the dynamic choice for workers to switch occupations in response to robots. To this end, I develop a fixed-point algorithm for solving the general equilibrium that features two-sided (firm and worker) heterogeneity and dynamics. I estimate that robots have increased average real wages by 0.8 percent but have lowered the real wages of production workers by 5.4 percent. Welfare losses from robots are concentrated on old production workers, as younger workers benefit from the option value of switching into tech.
Artificial Intelligence and College Majors, with Bjørn Meyer
Abstract: We study how Artificial Intelligence (AI) technology relates to the returns to college majors. Using microdata from Denmark, we first rank college majors according to whether their graduates work in AI firms. We show that AI cuts through the category of STEM degrees: while computer science and mathematics majors specialize in AI producer firms, we find that the laboratory sciences concentrate in firms that only use AI. We document that AI producer relevance correlates with higher wages, and that these earnings premiums are rising. In contrast, AI user majors do not earn wage premiums, nor have experienced differential earnings growth. Using a regression discontinuity design, we estimate the causal effects of admitting students to more AI-relevant college majors. We find that admission cutoffs are effective in increasing the AI relevance of college graduates. The causal earnings gain from higher AI producer relevance is at least as large as the correlation.
Works in Progress
Integrators: The Firm Boundaries of Capital-Skill Complementarity, with Gert Bijnens and Stijn Vanormelingen
Digitization of the Establishment-level Census of Manufactures from 1850, 1860, 1870, and 1880, with Richard Hornbeck and Martin Rotemberg
Equilibrium Impacts of Training Subsidies
Other Publications
Globalization, Flexicurity and Adult Vocational Training in Denmark (October 2019), with Jakob R. Munch Published in World Trade Organization book Making Globalization More Inclusive: Lessons from Experience with Adjustment Policies.