Measuring Statistical, Taste-Based, and Belief-Based Discrimination in AI and Human Decision Making: Evidence from Technical Recruitingwith Peter Bergman and Tushar Kundu
Artificial intelligence systems increasingly make consequential economic decisions. Generative AI has accelerated this trend: its pre-trained models can be quickly deployed in a variety of contexts. However, complex, proprietary models make it difficult to interpret large language model (LLM) decision making and to assess its biases relative to human decision makers.
We collect data on human and LLM hiring decisions for thousands of software-engineering candidates in a way that allows us to overcome the selection issue of observing hiring decisions only for interviewed applicants. The data allow us to readily assess accuracy and bias in evaluations.
We find that humans and the LLM are similarly accurate in their resume screening decisions on average, but the LLM is more selective. In terms of disparate impact, humans tend to recommend Black and Hispanic applicants at higher rates than white applicants after controlling for hiring outcomes, whereas the LLM produces substantial disadvantages for Black and URM applicants and a pronounced advantage for South Asian applicants.
We show that a model of applicant selection approximates both human and LLM decision making, allowing us to decompose bias into different forms of discrimination: taste-based discrimination, accurate statistical discrimination, and biased beliefs. We find that discrimination by human recruiters and the LLM reflects a combination of biased beliefs and taste-based discrimination, with taste-based discrimination playing a larger role for human recruiters.
Meet Kadeem Noray
Kadeem is an Assistant Professor at the Harvard Business School and Research Associate at Opportunity Insights at Harvard University. Kadeem's research studies talent selection and talent allocation in the economy. Kadeem is also interested in labor economics and public economics more broadly.