Research
(with Russell Legate-Yang)
September 2026
▷ Abstract
We present a robot exposure index based on how well robots can perform job tasks today. Robots can perform three-quarters of physical tasks in the US, making up 34% of working hours, but mostly in limited settings. Overall, about 80% of job tasks by working time are exposed to either robots or LLMs. The remaining unexposed work is highly interpersonal or requires physical skills that robots today don't have. Workers exposed to robots are more likely to be male, less educated, and lower paid. Driving and warehouse jobs are highly exposed to currently available robots; nursing and general repair jobs are not, since robots can do little of their work even in highly controlled environments. But while robots can do most physical work tasks today, they are much more expensive than human labor. Robots are cost-competitive for just 0.3% of job tasks. If robot price declines follow past trends, it will take 40 years for that share to reach 10%.
Robot exposure data
(with David Roodman)
August 2026
▷ Abstract
Could job retraining programs help address labor market disruption caused by AI? We analyze 56 randomized US studies through meta-analysis, finding that training programs yield positive but modest effects: for each person offered a training slot, employment rises by two to three percentage points. We identify a subset of "sector programs"—which partner with employers in high-demand fields and directly place workers—as producing significantly larger gains, though replication efforts have frequently failed. We conclude that existing programs would likely prove insufficient if AI causes large-scale worker displacement. Our key recommendation involves investing now in demonstrating, evaluating, and scaling promising initiatives, potentially through rapid expansion of leading programs with rigorous measurement.
(with Peter McCrory)
March 2026
▷ Abstract
We introduce "observed exposure," a measurement framework that combines theoretical LLM capabilities with real-world usage patterns to assess AI's labor market effects. Early analysis finds limited evidence of employment disruption to date, with some suggestive signs for younger workers in highly-exposed occupations.
(with Ruth Appel, Peter McCrory, Miles McCain, Ryan Heller, Tyler Neylon, and Alex Tamkin)
January 2026
▷ Abstract
This report introduces new metrics of AI usage to provide a rich portrait of interactions with Claude in November 2025, just prior to the release of Opus 4.5. These "primitives"—simple, foundational measures of how Claude is used, which we generate by asking Claude specific questions about anonymized Claude.ai and first-party (1P) API transcripts—cover five dimensions relevant to AI's economic impact: user and AI skills, how complex tasks are, the degree of autonomy afforded to Claude, how successful Claude is, and whether Claude is used for personal, educational, or work purposes.
(with Peter DeScioli, Kyle Thomas, and Steven Pinker)
Journal of Economic Behavior & Organization, November 2025
▷ Abstract
Why do people speak vaguely when they propose illicit deals? We examine the strategic use of vague speech. Participants play an economic game in which a schemer and accomplice can coordinate to take money from a mark. When a cop was watching, the schemer was more likely to send a vague message (``Some things are better left unsaid'') to the accomplice, which usually recruited the accomplice to collude. In Experiment 2, the schemer could write their own message. When the cop was watching, they wrote messages that were more vague, which again recruited the accomplice effectively.
(with Nathan Wilmers)
Journal of Public Economics, April 2025
▷ Abstract
We use location data to study activity and encounters across class lines. Low-income and especially high-income individuals are socially isolated: more likely than other income groups to encounter people from their own social class. Using simple counterfactual exercises, we study the causes. While some industries cater mainly to low or high-income groups (for example, golf courses and wineries), industry alone explains only a small share of isolation. People are most isolated when they are close to home, and the tendency to go to nearby locations explains about one-third of isolation. Using our uniquely detailed data, we show that brands, combined with distance, explain about half the isolation of the rich. Casual restaurant chains, like Olive Garden and Applebee's, have the largest positive impact on cross-class encounters through both scale and their diversity of visitors. Dollar stores and local pharmacies like CVS deepen isolation. Among publicly-funded spaces, libraries and parks are more redistributive than museums and historical sites. And, despite prominent restrictions on chain stores in some large US cities, chains are more class diverse than independent stores. The mix of establishments in a neighborhood is strongly associated with cross-class Facebook friendships (Chetty et al., 2022). The results uncover how policies that support certain public and private spaces might impact the connections that form across class divides.
Coverage: Washington Post, Bloomberg, The Economist, Planet Money, Marketplace, New York Times, Newsweek
October 2025
▷ Abstract
Is it feasible to alert workers of impending job destruction? For 80 years, the US Bureau of Labor Statistics has tried, providing thousands of employment forecasts in its Occupational Outlook. I test accuracy by comparing these forecasts to actual job growth. The forecasts were informative: across 4,000 unique predictions, occupations in the top third of projections grew significantly more in the subsequent three decades compared to ones in the bottom third. The forecasts were especially accurate when they made reference to business practices, like how a shift to cafeteria-style restaurants decreased the demand for waiters. Technology-focused forecasts were less accurate because, on average, they underestimated the extent of job loss. While the forecasts outperform naive extrapolations, the advantage is small, indicating that most of the explanatory power is captured in pre-existing trends. Overall, these results suggest that---at least historically---occupational growth and collapse has been foreseeable.
Coverage: Marginal Revolution
(with Kabir Dasgupta, Andrew C. Johnston, Linda Kirkpatrick, and Alexander Plum)
May 2025
NBER Working Paper No. 33873
▷ Abstract
How does family breakdown and divorce affect spouses and their children? We provide new evidence using a matched difference-in-differences design in rich administrative data from New Zealand. While most outcomes remain stable prior to separation, parents' mental health deteriorates in the lead-up. At separation, men's employment falls while women's rises, and women become much more likely to receive government benefits. Men temporarily double their criminal offending; about a third of the increase is domestic disputes. Both parents become more likely to be the victim of non-domestic crime as well. As for mental health, parents become more anxious and depressed at separation, and these remain elevated well after the couple has parted. Their children, too, face increased risks after separation: anxiety, depression, school absenteeism, and crime victimization all rise.
(with Evan Rose)
American Economic Journal: Applied Economics, 2024
NBER Working Paper No. 30385
▷ Abstract
We use administrative data from Washington State to perform a large-scale analysis of the impact of family formation on crime. Our estimates indicate that pregnancy triggers sharp declines in arrests rivaling any known intervention, supporting the view that childbirth is a "turning point" that reduces deviant behavior through social bonds. For mothers, criminal arrests drop precipitously in the first few months of pregnancy, stabilizing at half of pre-pregnancy levels three years after birth. Men show a sustained 20 percent decline in crime that begins at pregnancy, although arrests for domestic violence spike at birth. These effects are concentrated among first-time parents, suggesting that a permanent change in preferences---rather than transitory time and budget shocks---may be responsible. A separate design using parents of stillborn children to estimate counterfactual arrest rates reinforces the main findings. Marriage, in contrast, is not associated with any sudden changes and marks the completion of a gradual 50 percent decline in arrests for both men and women.
Coverage: Marginal Revolution, Mother Jones, The Economist
(with Aaron Chalfin and Aartik Sarma)
JAMA Surgery, 2024
▷ Abstract
In 2015, for the first time in nearly forty years, the rate of motor vehicle fatalities for Black Americans exceeded that of white Americans. By 2020, the gap in death rates stood at 34%, accounting for approximately 4,000 excess deaths between 2014 and 2020. This disproportionate increase occurred in nearly all states, in rural as well as urban areas, and was shared by drivers of all ages and genders. We consider a variety of potential explanations for the emerging race gap including race-specific changes in time spent driving, the circumstances of driving, the quality of medical care for crash victims, decreases in other types of mortality, changes in policing, and risky driving behaviors such as speeding, driving without a seat belt and driving while intoxicated. We can rule out many of these factors as important contributors to the race gap, but find evidence for two of them. The first is opportunity: Relative to white Americans, Black Americans are spending more time in vehicles than they have in the past. Changes in time spent driving, while modest, likely explain an important share of the emergent race gap. The second is a relative increase in drug use, manifested by a quadrupling of the rate of overdose deaths among Black Americans after 2014. Increased drug use appears to have resulted in a concomitant increase in fatal crashes involving drivers under the influence of drugs. Finally, we consider whether the emerging race gap is explained by the so-called "Ferguson effect," the idea that police officers have pulled back from enforcement activity in recent years. On the one hand, traffic stops made by police officers do appear to have declined after 2014. However, the decline in traffic stops does not appear to be race-specific and there is little evidence of a broad increase in risky driving behaviors like speeding and driving without a seat belt.
NBER version | Replication kit
(with Nathan Wilmers)
Journal of Labor Economics, 2024
▷ Abstract
From 1970 to 2000, US worker participation in labor strikes decreased by 90 percent. We show using understudied measures from labor market surveys that strikers also experienced worse outcomes after 1981. Event study evidence using the PSID suggests that strikers enjoyed 5-10 percent wage gains prior to the 1980s, but flat wage changes thereafter. Additional analysis of collective bargaining agreements and person-level data from the SIPP and CPS reinforce the finding that strikes since the 1980s have not been associated with increases in wages, hours, or benefits. These findings are consistent with decreased strike effectiveness, perhaps due to employers' higher propensity to hire strike replacements, or with more negative selection into "defensive" strikes that do not allow large pay increases.
Replication kit | Working paper
(with Nathan Wilmers)
American Economic Journal: Applied Economics, 2023
Best Paper Award, Academy of Management OMT Division, 2020.
▷ Abstract
Using new establishment-by-occupation microdata, we show that the use of discretionary wage-setting significantly expanded in the 1970s and 1980s. Increasingly, wages for blue-collar workers were not standardized by job title or seniority, but instead subject to managerial discretion. When establishments abandoned standardized pay rates, wages fell, particularly for the lowest-paid workers in a job and for those in establishments that previously paid above market rates. This shift away from standardized pay rates, in context of a broader decline in worker bargaining power, accelerated the decline in real wages experienced by blue collar workers in the 1980s.
Online Appendix | AEA Research Highlight | Replication kit
(with Daphné Skandalis and Ioana Marinescu)
2023
NBER Working Paper No. 30252 | Summary at Washington Center for Equitable Growth
▷ Abstract
The U.S. unemployment insurance (UI) system operates as a federal-state partnership, where states have considerable autonomy to decide on specific UI rules. This has allowed for systematically stricter rules in states with a larger Black population. We study how these differences in state rules create a gap in the unemployment insurance that Black and White unemployed workers receive. Using administrative data from random audits on UI claims in all states, we first document a large racial gap in the UI that unemployed workers receive after filing a new claim. Black claimants receive an 18% lower replacement rate (i.e., benefits relative to prior wage, including denials) than White claimants. In principle, the replacement rate of each claimant mechanically depends on the rules prevailing in her state and on her work history (e.g., the earnings before job loss and the reason for separation from prior employer). Since we observe claimants' UI-relevant work history and state, we are in a unique position to identify the role of each factor. After accounting for Black-White differences in work history, differences in rules across states create an 8% Black-White gap in replacement rate (i.e., slightly less than half of the overall gap). Using a standard welfare calculation, we show that states with the largest shares of Black workers would gain the most from having more generous UI rules. Altogether, our results highlight that disparate state rules in the UI institution create racial inequality without maximizing overall welfare.
2023
▷ Abstract
I document a surprising resilience to air pollution in the world's most polluted places, at least as measured by birth outcomes. Average birth weight and the incidence of low birth weight in some of the most polluted cities in the world, and in highly polluted US counties in the 1970s, are essentially identical to the contemporary US. This is puzzling since quasi-experimental studies find large negative impacts of air pollution on fetal health. I discuss several possible explanations.
Coverage: Marginal Revolution, Kevin Drum
(with Aaron Chalfin)
Proceedings of the National Academy of Sciences, 2022
▷ Abstract
This paper argues that changes in human activity during the COVID-19 pandemic led to an unusual divergence between crime rates and victimization risk in US cities. Most violent crimes declined during the pandemic. But analysis using foot traffic data shows that the risk of street crime victimization was elevated throughout 2020. People in public spaces were 15-30 percent more likely to be robbed or assaulted. This increase is unlikely to be explained by changes in crime reporting or selection into outdoor activities by potential victims. Traditional crime rates may present a misleading view of recent changes in public safety.
Coverage: Washington Post, Slate
Working paper | Replication kit
(with Julian De Freitas, Peter DeScioli, Jason Nemirow, and Steven Pinker)
Journal of Experimental Psychology: Learning, Memory, and Cognition, 2017
▷ Abstract
What is the relationship between the language people use to describe an event and their moral judgments?
We test the hypothesis that moral judgment and causative verbs rely on the same underlying mental
model of people's actions. Experiment 1a finds that participants choose different verbs to describe the
major variants of a moral dilemma, the trolley problem, mirroring differences in their wrongness
judgments: they described direct harm with a single causative verb (Adam killed the man), and indirect
harm with an intransitive verb in a periphrastic construction (Adam caused the man to die). Experiments
1b and 2 separate physical causality from moral valuation by varying whether the victim is a person or
animal and whether the harmful action rescues people or inanimate objects. The results show that
people's moral judgments lead them to portray a causal event as either more or less direct and intended,
which in turn shapes their verb choices. Experiment 3 finds the same basic asymmetry in verb usage in
a production task in which participants freely described what happened.
(with Peter DeScioli, Alex Shaw, Michael Bang Petersen, and Robert Kurzban)
Proceedings of the Royal Society B: Biological Sciences, 2014
▷ Abstract
Previous research emphasizes people's dispositions as a source of differences
in moral views. We investigate another source of moral disagreement,
self-interest. In three experiments, participants played a simple economic
game in which one player divides money with a partner according to the
principle of equality (same payoffs) or the principle of equity (payoffs
proportional to effort expended). We find, first, that people's moral judgment
of an allocation rule depends on their role in the game. People not only prefer
the rule that most benefits them but also judge it to be more fair and moral.
Second, we find that participants' views about equality and equity change
in a matter of minutes as they learn where their interests lie. Finally, we find
limits to self-interest: when the justification for equity is removed, participants
no longer show strategic advocacy of the unequal division. We discuss
implications for understanding oral debate and disagreement.
(with Michael Gurven, Christopher von Rueden, Hillard Kaplan, and Marino Lero Vie)
Journal of Personality and Social Psychology, 2013
▷ Abstract
The five-factor model (FFM) of personality variation has been replicated across a range of human
societies, suggesting the FFM is a human universal. However, most studies of the FFM have been
restricted to literate, urban populations, which are uncharacteristic of the majority of human evolutionary
history. We present the first test of the FFM in a largely illiterate, indigenous society. Tsimane
forager– horticulturalist men and women of Bolivia (n = 632) completed a translation of the 44-item Big
Five Inventory (Benet-Martínez & John, 1998), a widely used metric of the FFM. We failed to find robust
support for the FFM, based on tests of (a) internal consistency of items expected to segregate into the Big
Five factors, (b) response stability of the Big Five, (c) external validity of the Big Five with respect to
observed behavior, (d) factor structure according to exploratory and confirmatory factor analysis, and (e)
similarity with a U.S. target structure based on Procrustes rotation analysis. Replication of the FFM was
not improved in a separate sample of Tsimane adults (n = 430), who evaluated their spouses on the Big
Five Inventory. Removal of reverse-scored items that may have elicited response biases produced factors
suggestive of Extraversion, Agreeableness, and Conscientiousness, but fit to the FFM remained poor.
Response styles may covary with exposure to education, but we found no better fit to the FFM among
Tsimane who speak Spanish or have attended school. We argue that Tsimane personality variation
displays 2 principal factors that may reflect socioecological characteristics common to small-scale
societies. We offer evolutionary perspectives on why the structure of personality variation may not be
invariant across human societies.
Selected work in progress
Impacts of Congestion Pricing in New York City
(with Jesse Rothstein and Reed Walker)
Pre-analysis plan | OSF Repository
Helping out the outsourced: A pre-registered evaluation of the New Jersey temporary worker bill of rights
(with Jesse Rothstein)
▷ Abstract
We pre-register an observational study of New Jersey's New Jersey's Temporary Workers' Bill of Rights, which requires that temporary workers are paid the same as their full-time counterparts. Because the data for studying the effects is currently unavailable, this presents an opportunity for a pre-registered synthetic control design. Typically, observational studies contain hidden analytical decisions, potentially affecting the reliability of the estimates. A pre-registered synthetic control can address this issue by pinning down researcher decisions before the event takes places. This document establishes the exact data, code, outcomes, time periods, inference procedure, and robustness checks that we will use to study the reform. When the final period of data is available, we will run the code and report the results.
Pre-analysis plan | OSF Repository