Publications

I theorize that candidate filters, a core technology in digital labor markets considered central to cost-effective matching, redistributes costs from the search-and-screening stage to the job-offer stage of hiring. Filters cause employers to focus on applicants who display common preferred characteristics. At the search-and-screening stage, filters reduce the number of applicants interviewed and alter the composition of the interviewed pool by removing both lower- and higher-quality candidates. At the job-offer stage, filters lead employers to make job offers to candidates with easily observable signals of quality and higher wage demands. Additionally, I theorize that if many firms target the same signals of quality, they may all pursue similar candidates. This increases bargaining power for those applicants, leading to more rejected job offers and higher wage demands. A randomized field experiment in a large online labor market confirms that filters redistribute hiring costs. The experiment shows that employers with access to filters interview 3.4% fewer applicants. Using a machine learning algorithm to proxy for unobservable-to-the-employer applicant quality, I show that filters cause employers to interview 28.9% fewer lower-quality applicants and 36.9% fewer high-quality applicants per job. Among employers who actively used filters, I find suggestive evidence of worse job-offer outcomes: offers are 8.9% more likely to be rejected, and contracted wages are 2.5% higher. I discuss the implications of this cost redistribution for organizations, labor markets, and strategy research.

Extending the opportunity-discovery perspective on Kirznerian entrepreneurship, we propose a general framework in which new businesses emerge from three distinct mechanisms situated at the individual-opportunity nexus: discovery, discernment, and exploitation. We propose observing opportunities in market exchanges and characterizing their profitability potential based on a component that is common to all observers and one that is specific to individual observers who vary in access to market information. Analysis of an online platform for freelance labor demonstrates our contributions in theory, measurement, and inference. In this context, discovery and exploitation mechanisms shape individuals’ entrepreneurial transitions from freelancer to founder.

We report the results of a field experiment in which treated employers could not observe the compensation history of their job applicants. Treated employers responded by evaluating more applicants, and evaluating those applicants more intensively. They also responded by changing what kind of workers they evaluated: treated employers evaluated workers with 5% lower past average wages and hired workers with 13% lower past average wages. Conditional upon bargaining, workers hired by treated employers struck better wage bargains for themselves.

Platform marketplaces can use several means to steer buyers to select sellers, including recommendations and guarantees. We report the results of an experiment in which a platform marketplace guaranteed selected sellers for a group of would-be buyers. Offering a guarantee did not increase sales overall, suggesting financial risk was not determinative for the marginal buyer. However, the intervention did cause buyers to preferentially contract with guaranteed sellers — not because of lower financial risk, but because buyers viewed the platform’s decision to guarantee as informative about relative seller quality.

We examine how firms engage with the big data capabilities of two-sided matching platforms. We argue that firms deal with the appropriation challenge these platforms pose through strategies of partial reliance: they use the platform’s AI-driven recommendations to identify an initial set of generally acceptable partners, then rely on their own internal capabilities to select the best firm-specific match. Using a regression discontinuity design on an online labor platform, we find that firms rely heavily on the platform’s recommendations when screening applicants but ignore them when deciding whom to hire.

Organizational hierarchies economize on scarce expertise by dividing cognitive tasks vertically: juniors search for, assemble, and classify evidence, whereas seniors ratify what it means. We study how Artificial Intelligence (AI), inserted at the bottom of such a hierarchy, reshapes this division of labor in organizational decision-making. Drawing on theories of professional work, we distinguish two constituents of expert judgement: diagnosis and inference. We argue that AI more readily augments diagnosis, which involves searching and classifying evidence, than inference, which involves reasoning from evidence to warranted conclusions, because juniors and seniors exercise different forms of agency when working with the same AI system. Juniors improve diagnostic performance by deferring rule selection and search agency to AI, while interpretive discretion and decisional authority remain concentrated in accountable seniors. Using textual analysis of 212,784 junior-drafted and senior-validated radiology reports, we find that AI broadens search, increasing the diagnostic topics identified by juniors and validated by seniors. Yet AI reduces semantic similarity between the reports’ conclusions, signaling greater senior effort to ratify juniors’ inference. AI thus narrows the hierarchical gap in what is noticed while widening the gap in what it is taken to mean, redistributing decision-making costs from search to evaluation and increasing the burden on hierarchical gatekeepers.

Generative AI (GenAI) image models produce high-quality images at near-zero cost, raising concerns that they will displace human artists. The dominant task-based view predicts displacement wherever GenAI can perform an artist’s tasks. We argue this technical ability is necessary but not sufficient, because that view ignores whether buyers can appropriate the value of GenAI output. Appropriating the value of a creative work requires exclusive rights, which copyright law withholds from purely AI-generated work. Non-commercial buyers do not need those rights, so GenAI can displace the artists they commission; commercial buyers do, so their demand for human artists persists. Because the same artists serve both groups of buyers, we further predict that displaced artists carry the effects of GenAI into the commercial segment, depressing wages even where demand persists. We test these predictions on Artmug.kr, a Korean art-commission platform that bans AI-generated work and requires buyers to declare a commercial or non-commercial purpose. After DALL·E 2’s April 2022 release, non-commercial illustration demand fell 40% relative to commercial illustration demand, which showed no discontinuous decline. Displaced artists stayed and shifted their applications toward commercial commissions, and the commercial posted-wage premium fell from 20% to nearly zero by mid-2025. Because applications per posting did not rise, we attribute the posted-wage convergence not to intensified competition but to artists’ degraded outside options and selection among remaining non-commercial buyers. Task-based measures thus miss part of GenAI’s effect on commercial artists, whose wages fell even as demand for their work continued.

Studying how firms respond to fast-moving technologies requires observing where they invest in science. Each of the field’s standard measures, such as patents, publications, or R&D expenditure, makes only aspects of that investment visible. Patents and publications record what a firm’s science produces, but only when the firm chooses to disclose it. R&D expenditure records how much a firm spends, but not which sciences it spends on. We treat hiring as a revealed measure of firms’ investment in scientific human capital. A firm cannot do science without hiring people who can do it, so the jobs it posts reveal where it is investing before that investment produces a paper or a patent. An open-weight large language model reads about three million technical job postings at 2,583 US public firms from 2017 to 2025, identifying the roles that require keeping up with a research literature and the scientific fields they draw on. The resulting firm-year panel measures both the rate of a firm’s investment, the share of its technical hiring that engages with research, and its direction, the mix of sciences that hiring draws on. A firm that hires toward a field publishes and patents more in that field the following year than its industry peers. The measure extends the study of corporate science to firms that neither publish nor patent, which account for much of the research hiring in sectors such as finance and utilities. Because organizational capabilities reside substantially in people, the measure also captures not the capabilities a firm already holds but those it is building, and because it reads what each role requires rather than what it is called, it separates jobs that occupation codes treat as identical.