Digital Health Wire

LLMs Still Struggle With Medical Misinformation

A Lancet Digital Health study evaluated 20 large language models with over three million medical prompts, revealing a high susceptibility to misinformation. Neutral prompts led to a 32% acceptance rate of false information, which rose to 46% when embedded in formal discharge notes but dropped to 9% for social‑media content. Most logical fallacies actually reduced acceptance, except appeals to authority (+2.9 pp) and slippery‑slope (+2.2 pp). Larger models performed safer, yet specialized medical models lagged behind general‑purpose counterparts.