From Chatbots to MythTok: A Narrative Review of LLMs as Health Information Mediators on Social Media

Large Language Models (LLMs) are increasingly used to appraise health claims on social media (SM) platforms such as TikTok, yet no consistent approach exists for evaluating how they judge clinical accuracy or misleadingness. This narrative review develops a conceptual understanding of how LLMs should be evaluated as health information mediators in SM environments, using TikTok as a use case. We searched Google Scholar for peer-reviewed journal articles and conference papers, published between January 2022 and January 2026. Findings indicate that LLM outputs may be perceived as authoritative even when clinical accuracy is not independently verified. Identified evaluation approaches include reference-standard alignment, safety stress testing, knowledge-augmented verification, and prompt robustness testing. TikTok-focused studies highlight platform-specific risks, where entertainment-driven formats and influencer credibility can amplify misleading claims. Overall, robust evaluation requires four dimensions: alignment with clinical guidelines, prompt robustness, harm resistance, and sensitivity to platform-specific affordances such as short-form and multimodal content.