Abstract
BACKGROUND: Racial and ethnic bias in large language models (LLMs) used for health care tasks is a growing concern, as it may contribute to health disparities. In response, LLM operators implemented safeguards against prompts that are overtly seeking certain biases. OBJECTIVE: This study aims to investigate a potential racial and ethnic bias among 4 popular LLMs: GPT-3.5-turbo (OpenAI), GPT-4 (OpenAI), Gemini-1.0-pro (Google), and Llama3-70b (Meta) in generating health care consumer-directed text in the absence of overtly biased queries. METHODS: In this cross-sectional study, the 4 LLMs were prompted to generate discharge instructions for patients with HIV. Each patient's encounter deidentified metadata including race/ethnicity as a variable was passed over in a table format through a prompt 4 times, altering only the race/ethnicity information (African American, Asian, Hispanic White, and non-Hispanic White) each time, while keeping all other information constant. The prompt requested the model to write discharge instructions for each encounter without explicitly mentioning race or ethnicity. The LLM-generated instructions were analyzed for sentiment, subjectivity, reading ease, and word frequency by race/ethnicity. RESULTS: The only observed statistically significant difference between race/ethnicity groups was found in entity count (GPT-4, df=42, P=.047). However, post hoc chi-square analysis for GPT-4's entity counts showed no significant pairwise differences among race/ethnicity categories after Bonferroni correction. CONCLUSIONS: A total of 4 LLMs were relatively invariant to race/ethnicity in terms of linguistic and readability measures. While our study used proxy linguistic and readability measures to investigate racial and ethnic bias among 4 LLM responses in a health care-related task, there is an urgent need to establish universally accepted standards for measuring bias in LLM-generated responses. Further studies are needed to validate these results and assess their implications.