LLMs Reproduce Racial Stereotypes in Text Annotation

A recent study published on arXiv (2603.13891) indicates that large language models (LLMs) consistently mirror racial stereotypes during automated text annotation. Analyzing 19 LLMs across two experiments with over 4 million annotation assessments, the research found that subtle identity indicators influenced results. In a name-based experiment involving 39 annotation tasks, texts linked to Black names were perceived as more aggressive by 18 out of 19 models and more gossipy by the same number. For Asian names, a 'bamboo-ceiling' effect emerged: 17 models viewed individuals as more intelligent, while 18 assessed them as less confident and sociable. Arab names prompted cognitive elevation but interpersonal devaluation, with all four minority groups rated as less self-disciplined. Additionally, a matched dialect experiment revealed that the same sentence received different judgments based on dialect. This study emphasizes the dangers of employing LLMs in content moderation, hiring, and academic research, highlighting the necessity for thorough auditing of AI tools in sensitive contexts.