Large Language Models Reproduce Racial Stereotypes When Used for Text Annotation

ArXi:2603.13891v1 Announce Type: cross Large language models (LLMs) are increasingly used for automated text annotation in tasks ranging from academic research to content moderation and hiring. Across 19 LLMs and two experiments totaling than 4M annotation judgments, we show that subtle identity cues embedded in text systematically bias annotation outcomes in ways that mirror racial stereotypes. In a names-based experiment spanning 39 annotation tasks, texts containing names associated with Black individuals are rated as aggressive by 18 of 19 models and gossipy by 18 of 19.