AI is more likely than humans to form biases when hiring

Key takeaways

• LLMs develop stereotypes faster and more severely than humans in hiring scenarios—not just from training data bias, but from learning patterns in limited experience data
• The exploration-exploitation trade-off that LLMs optimize for makes them prone to premature generalization: one bad hire from an ethnic group triggers systematic exclusion
• Advanced reasoning models like o3 show worse bias outcomes (1.83 vs 0.84), suggesting capability scaling may amplify rather than mitigate stereotype formation
• As agentic AI systems gain memory and learning capabilities, they accumulate bias ammunition over time—a compounding risk for deployed hiring systems
• This research directly contradicts the 'AI removes human bias' narrative and has immediate regulatory/compliance implications for enterprise HR tech adoption