The spread of unverified health claims about drugs, dietary supplements, and alternative remedies on social media poses a growing public health concern. In this study, we present a retrieval-augmented generation (RAG) pipeline that uses large language models (LLMs) grounded in biomedical evidence from PubMed, openFDA adverse event reports, and NIH/NCCIH dietary supplement fact sheets to detect and classify health product misinformation. A total of 3493 health-related posts were collected from Reddit (948 posts across 12 subreddits) and YouTube (2545 video descriptions and comments), from which 8250 structured claims were extracted using Claude Haiku. Each claim was matched to biomedical evidence from three authoritative sources, achieving 79.4% evidence coverage, and classified into one of five veracity categories: supported (7.0%), unsupported (59.9%), exaggerated (22.4%), contradicted (2.0%), or dangerous (8.6%), together with an associated risk tier. Overall, 13.5% of claims were assigned high or critical risk. Cross-platform analysis showed that YouTube contained higher proportions of dangerous (11.3% vs. 2.9%) and exaggerated (27.0% vs. 12.4%) claims than Reddit. Compared with keyword-based and zero-shot transformer baselines, the LLM+RAG pipeline produced a more balanced and fine-grained classification of unsupported, exaggerated, contradicted, and dangerous claims. The most frequently implicated products were ashwagandha, kratom, black seed oil, turmeric, and ivermectin, with disease cure claims showing the highest dangerous classification rate (30.1%). These model-assigned results suggest that evidence-grounded LLM pipelines can support health misinformation surveillance, while also highlighting the need for expert validation and broader cross-platform evaluation.
Behnamian et al. (2026) studied this question.