Research from Princeton and the University of Chicago indicates that LLMs develop and act on biases more aggressively than humans when tasked with simulated hiring decisions.

Researchers from Princeton University and the University of Chicago found that large language models (LLMs) like ChatGPT, Claude, and Gemini exhibit higher levels of stereotyping than humans in simulated hiring experiments. The models, which were tasked with filling job roles from fictional ethnic groups, scored approximately 65% higher on a segregation scale than human participants. The study suggests that the same reasoning capabilities that allow AI to solve logic puzzles also drive them to overgeneralize from limited data.

Experiment Methodology and Findings

In the study, models acted as consultants for a fictional city, hiring candidates for 20 different roles. Candidates belonged to four fictional ethnic groups: Tufa, Aima, Reku, and Weki. Despite all candidates having an equal probability of success, the models quickly began segregating groups into specific job niches based on early, limited hiring outcomes. High-reasoning models, specifically OpenAI’s o3 and DeepSeek’s R1, demonstrated the most significant bias.

Addressing AI Bias

Researchers found that instructing models to be fair had little effect on their behavior. However, providing a bonus for diverse hiring successfully reduced bias. Additionally, models showed less reliance on ethnic stereotyping when provided with relevant personal information, such as age and education, rather than irrelevant traits like hair color. Experts note that as AI systems gain memory features, they risk locking in these learned biases, posing potential challenges for real-world applications like résumé screening.