Global Sentiment Toward Health AI at the Dawn of the ChatGPT Era: Empirical Analysis of Twitter (X) Discourse

Background: Artificial intelligence (AI) is increasingly proposed for use in health and health care systems. Beyond technical performance, public perceptions and affective responses influence whether AI technologies are accepted and adopted in real-world contexts. Social media platforms such as X (formerly Twitter) provide large-scale, real-time insight into public discourse surrounding emerging technologies, yet remain underused for examining how health AI is discussed, evaluated, and emotionally framed. Objective: This study aimed to develop and apply large language model (LLM)–based methods for exploratory social listening on health AI. This is the first study to map large-scale sentiment, emotional expressions, and confidence-related signals in online discussions of applications of AI to health. Methods: We collected 786,750 English-language posts from X (Twitter) published between January 1 and December 5, 2023, using health- and AI-related keywords. We benchmarked an LLM-based annotation framework by using OpenAI’s GPT-3.5-Turbo and GPT-4, comparing model classifications with trained human researchers. Annotations included overall sentiment and 6 evaluative domains frequently referenced in the literature surrounding attitudes toward health AI—usefulness, safety, privacy, ethics, quality, and trust. After cleaning, GPT-3.5-Turbo used the best-performing prompts to label 388,009 posts. A subset (n=268,347) was further analyzed using Emollama-7b, an open-source model fine-tuned from Meta’s LLaMA2-7B, for emotion detection, and latent Dirichlet allocation for thematic analysis. Comparisons were made across World Health Organization regions. Results: Compared against human annotations, optimized prompts achieved weighted F₁-scores above 0.60 across evaluative domains and sentiment classification. Global discourse about health AI was 65.26% (95% CI 65.11%-65.4%) positive and 83.62% (95% CI 83.48%-83.76%) emotionally optimistic, although substantial regional variation was observed in sentiment (P<.001). The Eastern Mediterranean and South-East Asia regions expressed significantly higher levels of positive sentiment and evaluative agreement in the studied features of health AI, alongside frequent discussion of the tech industry and commercial development. In comparison, the Western Pacific region expressed lower confidence and significantly more mentions of research topics (19.27%, 95% CI 18.5%-20.07%). Privacy was the most prominent global concern, with 33.31% (95% CI 32.98%-33.66%) of privacy-related posts expressing perceived risks. In the Region of the Americas, 18.19% (95% CI 17.92%-18.44%) of posts discussed algorithms and data governance, significantly higher than overall. Conclusions: This study offers the first systematic characterization of online health AI discourse at scale, mapping stances toward key features of AI, emotional tone, and discussion topics across regions. LLM-powered social listening is demonstrated as a feasible approach for identifying dominant narratives and regionally distinct concerns, capable of surfacing opinions absent from traditional media. This can extend to studying discourse on other evolving health technologies where public surveying is limited. While methodological refinement and multilingual expansion are needed, this framework can inform timely policy development, risk communication, and responsible health AI governance. Trial Registration: