- AI models, including advanced ones like OpenAI's o3, develop biases in hiring simulations, stereotyping job applicants more than humans do.
- In experiments, AI quickly generalized from limited data, segregating fictional ethnic groups into specific job roles based on early outcomes, even when all candidates were equally capable.
- The tendency to stereotype stems from LLMs' optimization for generalizing from few examples, which works in logical tasks but leads to bias in social settings.
- Newer models with higher reasoning capabilities showed stronger biases, raising concerns as chatbots gain memory features that may reinforce these stereotypes.
- Promising bonuses for diverse hiring reduced bias, suggesting that designing goals with social values can make AI act more desirably.
- Providing relevant personal information about individuals, such as age and education, also decreased segregation, while irrelevant details like hair color increased it.
- Real-world implications are uncertain since AI in hiring doesn't receive instant feedback on hire success, unlike in simulations.
- Researchers warn that AI could develop novel biases beyond those from training data, as it learns from experience in areas like hiring and loans.