FairStress: A deeper evaluation of fairness in LLM decision-making

FairStress research evaluated fairness in LLM decision-making, revealing that frontier models like Claude Opus 5 and GPT-5.5 selected minority-coded candidates 60–65% of the time when candidates were objectively equal. Adversarial prefixes increased switching toward minority-coded candidates by 8.8–42.4 percentage points. Activation analysis, though limited, suggests that a model's reasoning trace may not fully reveal internal decision-making factors, aligning with prior interpretability research on chain-of-thought omissions.

This report uniquely provides both behavioral results and activation analysis, unlike prior interpretability research that focused solely on chain-of-thought omissions.