AI Models Show Higher Bias Than Humans in Hiring Simulations

Researchers from Princeton University and the University of Chicago have found that large language models (LLMs) are more prone to developing biases and stereotyping job applicants than humans. In a simulated hiring game, LLMs like ChatGPT, Claude, and Gemini were tasked with filling 20 positions across various roles, including doctors, lawyers, and janitors, for a fictional city. Candidates were presented from four distinct ethnic groups: Tufa, Aima, Reku, and Weki. The AI models were informed about the success or failure of their hires in each round, with the objective of maximizing successful placements over 40 rounds. Crucially, all candidates were designed to have an equal probability of succeeding in any given role.

Following the initial rounds, the AI models began to exhibit discriminatory hiring patterns. They started to segregate candidates from different ethnic groups into specific job categories based on early hiring outcomes. For instance, if a model observed an Aima candidate failing in a doctor role, a position requiring high warmth and competence, it would subsequently avoid hiring Aimas for doctor positions. Instead, the models would disproportionately assign Aimas to janitor roles, which the AI classified as demanding lower levels of warmth and competence.

This tendency to stereotype based on demographic group was observed to be more pronounced in the AI models compared to human participants in the original psychology study that inspired the simulation. The research suggests that as AI companies develop more agentic models capable of remembering detailed user information, these models may accumulate data that fuels the formation of such biases. The study highlights a critical challenge in deploying AI for hiring processes, as the technology, intended to be objective, can inadvertently perpetuate and even amplify societal stereotypes.