Psychological Testing Reveals Racial Sentiment Biases in GPT-4 and GPT-3.5
July 31, 2026
Applying Implicit Association Test adaptations to LLMs, researchers found a small main effect of racial condition on sentiment scores across 126 prompts. The study evaluated GPT-3.5T, GPT-4, and GPT-4T to investigate how racial categories influence open-ended text sentiment.
HOW THIS AFFECTS YOU
●
researcherPsychological frameworks can be used to quantify unobservable mental processes and biases in LLM decision-making.
●
policyThese findings highlight the need for rigorous bias auditing in models deployed for government or healthcare.