AI Alignment Amplifies the Role of Race, Gender, and Disability in Hiring Decisions
创建于 更新于
摘要
Across 29 language models and 177 occupations (covering nearly half of U.S. employment), post-training alignment amplifies demographic effects in hiring: female and Black candidates gain advantages while disabled candidates face larger penalties; these effects rival several months to a year of education. Pre-trained models show smaller effects; alignment substantially increases sensitivity to credentials and shifts demographics in model representations, with disability underrepresented in training data and negatively encoded in deeper layers [page::0].
