LLMs stereotype job candidates roughly 65% more aggressively than humans do. In a Princeton and University of Chicago study published at ICML in Seoul, ChatGPT, Claude, and Gemini played a simulated hiring game across 40 rounds with fictional candidates from four ethnic groups. Human participants scored 0.84 on the study's segregation scale, where 2 is total group confinement. OpenAI's o3 scored 1.83.
The reasoning capability that makes models better at logic makes them worse at fairness. Higher-reasoning models, specifically o3 and DeepSeek's R1, showed the strongest stereotyping. The models are optimized to generalize from limited data, and in social contexts that optimization misfires fast. Telling models to 'be fair' barely moved results. Offering a bonus for diverse hiring did. The fix is in objective design, not value statements. The paper also found that relevant personal details about candidates reduced ethnic sorting, but irrelevant personal details did not.
The memory angle is the part worth reading in full. As chatbots gain persistent memory features, early biased observations can compound across sessions, and the researchers frame this as a direct risk for deployed hiring tools. The real-world feedback loop is slower than the experiment, but it still exists. Companies using LLMs to screen resumes or conduct interviews are building systems that learn from outcomes. Whether that learning reinforces bias over time is still an open question, and this paper is the clearest evidence yet that the risk is structural, not incidental.
[READ ORIGINAL →]