US Researchers 'Confirm Stronger Bias in AI Than Humans in Hiring Simulations'

AI has been found to exhibit significantly stronger hiring biases than humans.

Researchers from Princeton University and the University of Chicago conducted virtual hiring experiments involving prominent large language models (AI technologies that understand and generate text), such as ChatGPT, Claude, and Gemini. The researchers had these models assume the role of consultants for a fictional city and play a virtual hiring game involving 20 professions, ranging from doctors and lawyers to childcare workers and janitors. Four fictional population groups—Tufa, Aima, Lecu, and Weki—were established as applicants to analyze the AI's hiring patterns.

The researchers controlled the experiment so that the AI remained unaware that the candidates' probabilities of job success were identical. Nevertheless, the AI quickly demonstrated behavior that isolated and assigned specific groups to specific jobs based on just a few initial hiring outcomes. This amounted to stereotyping an entire group into different occupations based solely on initial observations that a few candidates from that group had failed in professional roles.

On an index measuring the extent of hiring segregation, human participants scored 0.84. In contrast, AI models showed a figure approximately 65% higher, revealing a tendency to segregate population groups more quickly and extremely than humans.

More advanced models with higher reasoning capabilities were found to be at greater risk of falling into bias. According to a paper presented at the International Conference on Machine Learning (ICML), OpenAI's 'o3' model, which boasts top-tier reasoning performance, recorded a hiring segregation score of 1.83.