AI agents outpace humans in creating hiring bias
a princeton study found that llms like gpt-4 and claude exhibit 65% more hiring bias than humans. this suggests that advanced reasoning capabilities and long-term memory could inadvertently automate discrimination.
researchers found that llms, including openai's o3 and deepseek's r1, are 65% more likely than humans to stereotype job applicants based on demographic traits. in a simulated hiring game, models quickly segregated fictional ethnic groups into specific job roles after observing just one success or failure, scoring 1.83 on a segregation scale compared to the human score of 0.84. this over-generalization stems from the models' optimization for tasks requiring pattern recognition, a trait that becomes dangerous when applied to social decision-making. as ai labs integrate long-term memory into agents to enable personalization, these models risk reinforcing past interactions and cementing biases.