[Seminaire CREM] Generative AI and Labor Market Discrimination

Authors: Kenza Elass (Bocconi University) Germain Gauthier (Bocconi University), Debora Nozza (Bocconi University) & Paola Profeta (Bocconi University)

Abstract:
With the rise of Large Language Models (LLMs) -- such as OpenAI's ChatGPT -- there has been a proliferation of websites offering CV generation services for job seekers, as well as CV screening services for employers. We study the use of LLMs to produce and screen CVs of job applicants and make wage recommendations. We run a series of online experiments inspired by classical correspondence studies. Four main findings emerge. First, leading open-source and proprietary models typically produce gender-stereotyped CVs. Second, they discriminate against women. Consistent with theories of systemic discrimination, the bias against women increases if the models are used to produce and screen applicant packages. Third, when making salary recommendations, LLMs consistently suggest lower earnings for women. Fourth, these effects are particularly pronounced in male-dominated or highly unequal occupations. Lastly, we also explore ways to mitigate these effects.