Research

We study the 2013 U.S. federal government shutdown and its impact on federally funded research in Antarctica. Although the shutdown lasted only 16 days, it coincided with the start of the Antarctic summer field season, a narrow window for data collection that relies on geographically unique, time-sensitive, and difficult-to-substitute research inputs. Using scientist-level data with a difference-in-differences design and exploiting the timing of the decision, we document an 10% decline in the number of by-line weighted publications among affected researchers, as well as altered collaboration patterns, which we corroborate with qualitative survey evidence. These findings show that even brief disruptions can have lasting effects on science when they block access to specialized, time-sensitive, and hard-to-replace research inputs.

Abstract removed for now!

Scientific discovery depends not only on what ideas are funded, but on how researchers translate ideas into resource requests that shape teams, training, and future knowledge production. Using a unique dataset covering the full universe of funded and unfunded grant applications to two major philanthropic science funders in Europe between 2011 and 2022, linked to peer-review scores, investigators’ prior research records, and itemized budgets, we examine where gender differences arise in the funding process. We document that female principal investigators receive 6.4–7.35% less funding per application than male investigators, a gap that is almost entirely explained differences at the proposal stage: women request 5-5.4% less funding, conditional on proposal quality and researcher characteristics, despite being overall more likely to receive funding. This is despite there being no incentives to request smaller grants, a claim we confirm empirically. Using granular budget data, we show that this gap is highly concentrated in team size formation, with female investigators systematically proposing smaller PhD student teams, with no comparable differences across other budget categories. Consistent with learning and information frictions as a mechanism, the gender gap in funding requests is concentrated among early-career researchers with weaker publication signals and lower institutional prestige, and it converges with experience through repeated applications. These findings shift attention from evaluators to applicants, highlighting how differences in resource construction - rather than evaluation bias alone - can generate durable disparities in scientific capacity. Extrapolating to the U.S. context suggests that women may be systematically requesting on the order of $0.75 billion less annually from major public funders.

Artificial intelligence is advancing at a pace once thought unimaginable, yet we still lack clear tools to understand how these breakthroughs map onto the world of work. This paper introduces a novel framework that systematically links AI benchmark progress - the scoreboards that track frontier capabilities - to the occupational tasks that define human labor. Unlike patents, surveys, or deployment data, which are often lagged, opaque, or subjective, benchmarks are transparent, replicable, and updated in near real time. Using O*NET as a bridge, we connect benchmark trajectories across domains-including language, reasoning, vision, and multimodal tasks-to 52 human abilities, and translate these into occupation-level indices of AI exposure. The result is a dynamic, task-level methodology that allows us to track and forecast where automation pressures are likely to emerge. By repositioning benchmarks from technical scoreboards to economic indicators, this study offers a fresh lens for anticipating the future of work and shaping policy responses.

This paper introduces WAGE-Bench, a new framework for evaluating artificial intelligence based on the practical economic value it creates in human work. Standard benchmarks measure technical capability but reveal little about how much AI actually reduces the cost of completing real tasks. WAGE-Bench addresses this gap by eliciting individuals’ willingness to accept compensation for performing realistic tasks under alternative AI conditions - without AI, with AI assistance, or with alternative models. Differences in willingness to accept yield a common monetary metric of AI’s practical value, enabling comparisons across models, tasks, occupations, and user populations. The framework occupies a middle ground between conventional benchmarks and randomized controlled trials: cheaper, faster, and more scalable than RCTs, yet more closely tied to real-world usefulness than standard benchmarks. We develop the framework conceptually and illustrate its feasibility through a pilot experiment in which undergraduate students value familiar tasks with and without AI access. The pilot reveals substantial variation in the perceived value of AI across tasks, with the largest gains concentrated in technical and cognitively demanding activities. WAGE-Bench thus offers a scalable and economically interpretable complement to existing evaluation methods, shifting AI evaluation from abstract performance toward practical usefulness. 

This paper studies how permanent agricultural productivity shocks shape the long-run incidence of conflict. We examine the introduction of the potato to the Old World, a major productivity shock from the Columbian Exchange, using a panel of 400 km × 400 km grid cells from 1400–1900 and historical battle data from Kitamura’s World Historical Battles Database. Exploiting spatial variation in potato suitability before and after its diffusion, we find that areas more suitable for potato cultivation experienced higher subsequent battle incidence. The effect is substantially larger in areas with higher lactase persistence, consistent with nutritional complementarities between potatoes and milk amplifying the demographic and economic value of land. The results suggest that permanent productivity gains can increase conflict when they raise land values, population pressure, and the returns to territorial control.