Research

How does Artificial Intelligence (AI) affect the task content of jobs, and how do workers adjust to AI’s diffusion? To answer these important questions, we combine novel patent-based measures of AI and robot exposure with individual survey data on job tasks and administrative data on worker careers. Like prior studies, we find that robots have reduced routine tasks. In sharp contrast, AI has reduced abstract tasks like information gathering and increased the demand for complex routine tasks like monitoring processes. These task shifts affect all skill groups, mainly occur within detailed occupations, and become stronger over time. While displacement effects are small, workers have responded by switching jobs, often to less exposed industries. The AI-related task shifts and negative wage effects are stronger for low-skilled workers and older incumbents. Labor market entrants, in turn, face lower entry wages and fewer employment opportunities in exposed industries.

This paper examines preferences for gender diversity among co-workers. Using stated-choice experiments with more than 9,200 professors, PhD students, and university students in Germany, we uncover a substantial willingness to pay (WTP) for gender diversity of up to 5% of earnings on average. Importantly, we find that women have a much higher WTP for gender diversity than men. While the WTP differs by career ambition and related characteristics like competitiveness and family preferences, we find that gender differences in these dimensions cannot explain the gender gap in the WTP for diversity. Our findings provide an explanation for differential sorting of men and women into high-profile jobs based on the share of female co-workers.

Worker perceptions of pay and key job characteristics are important determinants of individual labor market behavior and aggregate labor market outcomes. In this paper, we characterize workers’ perceptions about the link between wages and amenities across jobs. To this end, we carry out a survey among more than 3,000 workers in Germany. On average, perceptions are in line with the existence of compensating differentials, i.e., a trade-off between wages and amenities. However, perceptions are vastly heterogeneous and differ across amenities and across workers for a given amenity. Workers subject to lower labor market frictions and those with recent job search activity are more likely to perceive a trade-off between wages and amenities. We find indirect evidence of considerable misperceptions, as workers are anchoring their beliefs about the market-level wage-amenity trade-off around their own personal valuation of amenities.

We use a generalized random forest to provide new insights into the link between exports and earnings inequality between workers. Using German administrative data and exploiting industry-level variation in exports to China and Eastern Europe over time, we estimate worker-level earnings responses to exports across skill, demographic, occupational, and firm dimensions. We find that the workers with the largest earnings gains are younger, more likely to be male, high-skilled, and employed in specialized occupations and in large firms. Demographic and occupational characteristics are as important as firm and skill attributes in explaining heterogeneity in the effects. The earnings gains peak around the 70th percentile of the pre-shock earnings distribution, indicating that export-induced inequality arises mainly in the lower half rather than at the top. Overall, the effects on inequality are modest, as the gains from exports are broadly shared across the earnings distribution.

We use causal machine learning to revisit the link between electronic monitoring (EM) and criminal recidivism. Combining a generalized random forest with the judge instrumental variable design by Di Tella and Schargrodsky (2013), we confirm a negative average effect of EM on rearrest rates. We uncover treatment effect heterogeneity linked to demographics and the criminal history, but the results indicate that no individuals experience an increased rearrest probability from EM. Our findings highlight the benefits of causal machine learning for policy-relevant questions and suggest that the effectiveness of EM depends on the composition of the target population.