AI is more likely than humans to form biases when hiring (www.technologyreview.com)

Recent research from Princeton University and the University of Chicago highlights a troubling aspect of AI hiring practices: large language models (LLMs), like ChatGPT and Claude, are more prone to forming biases when making hiring decisions than humans are. In an experiment simulating a hiring process, these models, tasked with hiring candidates from four fictional ethnic groups, developed stereotypes based on initial hiring outcomes. They demonstrated a significant tendency to segregate candidates by ethnicity into job roles, with LLMs scoring 1.83 on a segregation scale compared to 0.84 for human participants, reflecting their rapid generalization from limited data. This study is particularly significant as AI becomes increasingly integrated into hiring processes, raising concerns about biases that may emerge from both training data and real-world experiences. The researchers found that instructing models to be fair had minimal impact, but encouraging diverse hiring goals significantly reduced bias. Notably, when given pertinent personal information rather than irrelevant data, the models were less likely to discriminate based on ethnicity. As AI continues to influence critical decisions in hiring, lending, and more, the potential for these novel biases necessitates careful consideration and the incorporation of social values into AI design to foster equity and fairness.