Large Language Models Generate Stigmatizing Language During Reasoning Over Real-World Clinical Data

Stigmatizing language in clinical documentation, which conveys negative stereotypes, attitudes, or judgments toward patients, is a recognized source of documentation bias and is associated with poorer care and adverse health outcomes. Although prior stigma-related research has explored on clinician-written EHR notes, the increasing use of large language model (LLM)-generated documentation in clinical workflows raises new concerns about its potential to produce or amplify bias and affect patient safety. In this study, we conducted a large-scale assessment of stigmatizing language in LLM-generated reasoning text on 35 real-world clinical tasks across 107 LLMs. We applied a psychiatrist-validated, natural language processing (NLP) system to detect stigma terms in LLM reasoning text and quantified stigma rates of LLM-generated reasoning texts across 3,745 model-task pairs. Results showed that stigma rates ranged from 0% to 33.33%, with 84.06% of pairs containing stigma terms. Reasoning models showed higher stigma rates than non-reasoning models (2.35% vs. 1.70%; p < 0.0001), whereas medical models did not show significantly lower stigma rates than general-purpose models (1.80% vs. 2.00%; p = 0.26). Stigma rates of LLM outputs correlated negatively with task accuracy (r = -0.283; p < 0.001) and positively with input clinical-text stigma (r = 0.569; p < 0.001), with 19.76% of LLM-task pairs amplifying stigma in the original input notes. The effectiveness of prompt engineering as a destigmatizing approach varied across models, with stigma rates reduced by up to 91.91% without compromising model performance. This study shows that stigmatizing language generation is common but modifiable in LLM-generated reasoning traces, highlighting the need for direct stigma-related safety evaluation and robust destigmatization strategies before clinical deployment.