AI is more likely than humans to form biases when hiring
AI can both inherit human stereotypes and generate new biases, with implications for fair hiring as models gain more memory and decision‑making power.
At a glance
- A Princeton-Chicago study simulated hiring with LLMs (including ChatGPT, Claude, Gemini) across 20 jobs and four demographic groups, learning from outcomes over 40 rounds.
- Models tended to segregate candidates by group, even though all candidates had equal success probabilities, with higher bias in newer, more capable models (e.g., OpenAI o3, DeepSeek R1).
- Humans scored 0.84 on a segregation scale, while models scored roughly 65% higher (o3 around 1.83, near the top of the scale).
- Promising better societal outcomes, the researchers found that explicitly rewarding diverse hiring reduced bias in models.
The story
MIT Technology Review summarizes a study where Princeton and University of Chicago researchers tested LLMs by having them act as hiring consultants for 20 jobs across four fictional ethnic groups. In each round, candidates had equal chances of success, but models tended to assign jobs based on early outcomes, effectively forming stereotypes about groups.
Newer models with greater reasoning capabilities, such as OpenAI’s o3 and DeepSeek’s R1, showed stronger biases than earlier versions. On a segregation scale, where 2 means complete separation, humans scored 0.84, while models scored roughly 65% higher, with o3 at about 1.83.
Researchers note that LLMs generalize from limited data because generalization is a core objective of their training. The study suggests that if models are incentivized to favor diverse hiring, biases can decrease.
The study also notes that providing more personal information about individuals can influence bias dynamics, and that while asking models to be fair did not dramatically shift behavior, adjusting the goal structure to include social value considerations can improve outcomes.