Safety, Accuracy, Empathy, Information Quality, and Readability of Chronic Obstructive Pulmonary Disease Health Advice Generated by 6 Large Language Models: Cross-Sectional Comparative Evaluation Study
Year: 2026
Type: article
Abstract: BACKGROUND Large language models (LLMs) are increasingly used for health information, but chronic obstructive pulmonary disease (COPD)–related questions often involve safety-sensitive decisions, including treatment escalation, oxygen therapy, and exacerbation management. Because generative outputs may vary across repeated prompts, response stability is also relevant to comparative evaluation. OBJECTIVE This study aims to compare 6 public-facing LLM chatbots in answering COPD-related public health questions across safety, accuracy, empathy, information quality, source transparency, readability, and 2-round response stability. METHODS We developed a 55-question English COPD question bank from an initial 108-question pool using web-based retrieval, Google Trends, PubMed-informed topic review, patient education resources, and expert screening. Six public-facing chatbot products were evaluated in a primary first round and a post hoc independent second round, added after first-round data collection, using the same 55 questions. Visible model/mode labels are provided in the supplementary materials. Five blinded raters assessed manually scored outcomes against a Global Initiative for Chronic Obstructive Lung Disease (GOLD) 2026-centered reference standard, and 6 readability indices were calculated. The primary cross-model analysis used the first round; the second round was analyzed as a matched retest sensitivity analysis. Stability was assessed using exact agreement and Cohen κ for safety, quadratic-weighted κ for ordinal ratings, and absolute-agreement single-measure intraclass correlation coefficients (ICCs) for continuous scores and readability indices; 95% CIs used question-level bootstrap resampling. RESULTS The first round yielded 330 responses. Unsafe responses occurred in every model: ChatGPT, Gemini, Copilot, and DeepSeek each generated 4 unsafe responses (7.3%); Qwen generated 5 (9.1%); and Doubao generated 6 (10.9%). The overall safety comparison was not statistically significant (Cochran Q=1.909, df=5; P=.86); this finding should not be interpreted as evidence of equivalent safety. Accuracy, empathy, information-quality measures, source-transparency scores, and readability indices differed significantly across models. Across 330 matched question-model pairs, safety classifications agreed in 97.3% of pairs (κ=0.821, 95% CI 0.543-0.979); accuracy and empathy showed weighted κ values of 0.872 and 0.856, respectively; DISCERN (an instrument for judging the quality of written consumer health information on treatment choices) and adapted Ensuring Quality Information for Patients (EQIP) showed ICCs of 0.954 and 0.925. Most readability indices also showed high overall 2-round stability, although model-specific variability remained. Nine safety classifications changed across the 2 rounds (5 from safe to unsafe and 4 from unsafe to safe). CONCLUSIONS In this standardized single-turn COPD question set, model performance differed across several evaluated outcomes, whereas the first-round safety test neither established a between-model difference nor demonstrated equivalence. All 6 models produced potentially unsafe advice, and 9 safety classifications changed between rounds conducted several months apart despite high pooled agreement. These findings support cautious use of LLMs for supplementary COPD education, with clinician oversight for individualized treatment decisions and ongoing reassessment as public-facing systems change.
Source: JMIR Medical Informatics
DOI: 10.2196/104836
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,300
OpenAlex ID: W7225644736