LLMs can form hiring biases from limited experience

Researchers at Princeton University and the University of Chicago found that LLMs can develop their own biases from experience, not only absorb human biases from training data. In a simulated hiring game involving ChatGPT, Claude, Gemini and other models, each system chose candidates for 20 jobs over 40 rounds, then learned whether each hire succeeded. All fictional ethnic groups were equally likely to succeed, but the models began assigning groups to different job niches after limited feedback.

The models stereotyped candidates more strongly than human participants in the original psychology experiment. On a segregation scale where 2 means every group is completely confined to its own niche, humans scored 0.84, while the models scored roughly 65% higher. OpenAI’s o3 reached 1.83, close to the maximum possible, and newer reasoning models such as o3 and DeepSeek’s R1 showed stronger biases.

Simple fairness instructions had little effect, but offering a bonus for diverse hiring reduced bias. The models also relied less on ethnicity when given relevant personal details, such as age and education, but reverted to ethnic sorting when given irrelevant details. The findings point to risks for AI systems that screen résumés, conduct interviews, or use memory and feedback in decisions about jobs, loans, or parole.