Evaluating Anti-LGBTQIA+ Medical Bias in Large Language Models

Abstract

From drafting responses to patient messages to clinical decision support to patient-facing educational chatbots, Large Language Models (LLMs) present many opportunities for use in clinical situations. In these applications, we must consider potential harms to minoritized groups through the propagation of medical misinformation or previously-held misconceptions. In this work, we evaluate the potential of LLMs to propagate anti-LGBTQIA+ medical bias and misinformation. We prompted 4 LLMs (Gemini 1.5 Flash, Claude 3 Haiku, GPT-4o, Stanford Medicine Secure GPT (GPT-4.0)) with a set of 38 prompts consisting of explicit questions and synthetic clinical notes created by medically trained reviewers and LGBTQIA+ health experts. The prompts explored clinical situations across two axes: (i) situations where historical bias has been observed vs. not observed, and (ii) situations where LGBTQIA+ identity is relevant to clinical care vs. not relevant. Medically trained reviewers evaluated LLM responses for appropriateness (safety, privacy, hallucination/accuracy, and bias) and clinical utility. We find that all 4 LLMs evaluated generated inappropriate responses to our prompt set. LLM performance is strongly hampered by learned anti-LGBTQIA+ bias and over-reliance on the mentioned conditions in prompts. Given these results, future work should focus on tailoring output formats according to stated use cases, decreasing sycophancy and reliance on extraneous information in the prompt, and improving accuracy and decreasing bias for LGBTQIA+ patients and care providers.