Website CV Paper
Fields
Labor, Behavioral & Experimental, Development, Applied Micro, Innovation
Job market paper
Humans, Generative AI, and Disparities in Access to Employment
How does delegating applicant screening to generative AI change who gets hired? We study an Italian job platform that screens applicants on behalf of 1,433 employers and, in 2025, moved screening from human reviewers to an LLM. We observe which candidates progress through each stage of the hiring process, from applications to shortlisting through a first round of screening, and finally the hiring decisions by the employer. We establish three results. First, human screeners discriminate: within the same vacancy, applicants with immigrant-sounding names are 27% less likely to be shortlisted than applicants who are otherwise identical on everything the screener observes on the platform. The penalty differs sharply and persistently across reviewers applying the same rubric, even for the same employer. Second, we compare human shortlisting decisions with AI evaluations of the same applications, holding constant the number of candidates. AI cuts the penalty by 63% and shrinks its dispersion across reviewers' applicant pools. Third, exploiting as good as random variation in the assignment of vacancies to different screening regimes, we provide causal evidence that shifting screening authority to AI, relative to reviewers who already see AI scores, closes the ethnic hiring gap and leads to faster shortlisting without reducing total hires or employer satisfaction. By contrast, the age penalty, which AI also removes at screening, reappears in employers' hiring decisions. AI screening can remove discrimination introduced by intermediaries, but not discrimination rooted in employers' own preferences.