A study published in The Lancet Digital Health in 2026 tested how readily large language models reproduce medical misinformation, and found that they often will. Researchers at the Icahn School of Medicine at Mount Sinai exposed nine leading models to more than one million prompts, drawing on three kinds of material: real hospital discharge summaries seeded with fabricated recommendations, health myths taken from social media, and 300 short clinical scenarios written and validated by physicians.

When a false statement was presented in the register of routine clinical documentation, the models tended to carry it forward rather than challenge it. A fabricated recommendation inserted into a discharge note could be restated as though it were established practice. "A fabricated recommendation in a discharge note can slip through. It can be repeated as if it were standard care," said Eyal Klang, a co-senior author and chief of generative AI at the institution.

The authors argue that susceptibility to misinformation should be treated as a measurable property of a medical AI system. They recommend large-scale stress tests before deployment, built-in checks against external evidence before a model presents a claim as fact, and re-testing with each new model generation. "Instead of assuming a model is safe, you can measure how often it passes on a lie," said the first author, Mahmud Omar. The dataset assembled for the study is offered as one such test.

The work measures behaviour on curated prompts rather than in live clinical use, so the rate at which such errors would reach patients in a deployed system remains open. Its contribution is a reproducible way to quantify the failure mode before that happens.