A new Lancet Digital Health study calls for built-in safeguards to verify medical claims before AI presents them as fact, aiming to curb medical misinformation.

Researchers tested 20 language models—both open-source and proprietary—across three content types: real hospital discharge notes with a fabricated recommendation, Reddit health myths, and 300 physician-written clinical scenarios.

The Icahn Mount Sinai study investigates whether medical AI can propagate misinformation when it mimics realistic clinical and social-media language.

OpenAI’s GPT models were the most accurate at detecting fallacies and resisting misinformation, while other models showed up to 63.6% susceptibility to false claims.

Funding for the study comes from CTSA grant UL1TR004419 and NIH S10 infrastructure awards S10OD026880 and S10OD030463.

Misinformation from Reddit or other social-media sources was propagated by AI far less often, around 9% of the time, indicating lower acceptance of casual-language myths.

Mount Sinai’s Windreich Department of AI and Human Health, in collaboration with the Hasso Plattner Institute for Digital Health, exemplifies the institution’s broader push for safe, ethical health AI initiatives.

An illustrative discharge-note example falsely advising drinking cold milk for esophagitis-related bleeding shows several models treated the claim as safe, revealing how confident medical language can be misread as true.

The paper, titled Mapping LLM Susceptibility to Medical Misinformation Across Clinical Notes and Social Media, appears in The Lancet Digital Health in February 2026, with Mahmud Omar et al. as authors and recognition of CTSA and NIH support.

Prompt framing matters: authoritative-sounding prompts increase the likelihood that models endorse false information.

Across over a million prompts, AI acceptance of fabricated information was about 32% overall, rising to nearly 47% when misinformation resembled authentic discharge notes, and dropping to 9% for Reddit-derived content.

Leaders emphasize that confident, well-written medical language can be treated as true by models, revealing a vulnerability that needs addressing.