This cross-sectional benchmarking analysis evaluates the susceptibility of nine large language models to medical misinformation. The study analyzes more than 1.7 million outputs generated from 1,000 emergency department cases presented in multiple sociodemographic variations while keeping the clinical content constant. The work examines whether models accept or propagate medically unjustified recommendations and how this varies across contexts such as clinical notes and social media. It is relevant for understanding robustness, misinformation sensitivity, and the safe deployment of LLMs in medical settings.

Author of the paper: Mahmud Omar, Vera Sorin, et al.

Publisher or journal of publication: The Lancet Digital Health

The paper is available at the following link.