Large Language Models Portray Socially Subordinate Groups as More Homogeneous, Consistent with a Bias Observed in Humans

Keyphrases

• Large Language Models 100%
• Chat Generative Pre-trained Transformer (ChatGPT) 50%
• African American 33%
• Hispanic Americans 33%
• White Americans 33%
• Asian Americans 33%
• Racial-ethnic Groups 16%
• Stereotypic Behavior 16%
• Intersectional 16%
• One-form 16%
• Training Data 16%
• Everyday Life 16%
• Social Groups 16%
• Racial Minority Groups 16%
• Psychological Phenomena 16%
• Social Psychological 16%
• Human Bias 16%
• Human Experience 16%
• Discriminatory Behavior 16%
• Group Identity 16%

Psychology

• White Americans 100%
• Asian Americans 50%