User evaluation of AI- and human-generated responses in digital health communication: a paired survey study

Abstract

Digital health communication is gaining importance as health care systems face increasing demand and structural constraints. Generative artificial intelligence (AI) tools such as ChatGPT are increasingly used for health information seeking; however, direct comparisons between AI-generated and human responses from the user perspective remain limited. A quantitative exploratory cross-sectional online survey using a paired-response design was conducted in Germany. The survey instrument was developed based on the study objectives and refined through pretesting. Participants evaluated two responses to the same real-world health query: one written by a physiotherapist and one generated by ChatGPT-4. Outcomes included perceived empathy, comprehensibility, perceived discriminatory content in the responses, and expert-rated clinical quality. Paired comparisons were analyzed using Wilcoxon signed-rank tests, and logistic regression models were applied to explore associations between participant characteristics and preference patterns. The analytical sample comprised 224 participants. Across most items, the AI-generated response was rated more favorably than the human response, particularly with regard to comprehensibility and empathy (all p < .001). The AI-generated response was rated more favorably across both empathy and comprehensibility items, with highly comparable preference distributions between the two domains. Perceptions of discriminatory content were rare and did not differ significantly between responses. Expert evaluations also favored the AI-generated response in terms of clinical accuracy and completeness, although these findings should be interpreted cautiously given the exploratory design. The findings suggest that participants perceived the AI-generated response as aligning more closely with central user expectations for text-based health information, particularly regarding clarity and supportive language, although these findings should be interpreted cautiously given the substantial differences in length and structure between responses and the exploratory single-case design. At the same time, these perceptions should be considered alongside ongoing concerns regarding the clinical appropriateness, safety, and equity of AI-generated health information, which were not comprehensively assessed in the present study. The present findings therefore primarily provide evidence regarding user-perceived communication quality rather than the clinical validity of AI-generated responses. Accordingly, further research is needed to establish the clinical appropriateness, safety, guideline concordance, and potential role of AI-generated responses in healthcare consultation.