AI is more likely than humans to form biases when hiring

A simulated hiring study found that leading language models quickly formed stereotypes from limited feedback, despite all fictional groups having equal chances of success. The models displayed more segregation than humans, but diversity rewards and relevant personal information reduced their bias.

一项模拟招聘研究发现，尽管所有虚构群体的成功概率相同，主流大语言模型仍会根据有限的反馈迅速形成刻板印象，而且表现出的群体隔离倾向比人类更强。不过，提供多元化招聘奖励和相关个人信息可以减轻这种偏见。