Large Language Models Reproduce Racial Stereotypes When Used for Text Annotation

A comprehensive study of 19 large language models reveals systematic racial bias in automated text annotation, with over 4 million judgments showing LLMs consistently reproduce harmful stereotypes based on names and dialect. The research demonstrates that AI models rate texts with Black-associated names as more aggressive and those written in African American Vernacular English as less professional and more toxic.

• →All 19 tested LLMs showed systematic racial bias when annotating text, with 18 of 19 models rating Black-associated names as more aggressive and gossipy.
• →Asian names triggered a 'bamboo ceiling' effect where 17 of 19 models rated individuals as more intelligent but 18 of 19 as less confident and sociable.
• →African American Vernacular English was consistently judged as less professional and more toxic compared to Standard American English across nearly all models.
• →All minority groups were rated as less self-disciplined, while Arab names elicited cognitive elevation alongside interpersonal devaluation.
• →These biases directly embed into datasets used for research, governance, and decision-making as LLMs become more widely adopted for automation.