Large language models exhibit greater diagnostic anchoring than physicians in a forced-choice vignette study.

Anchoring bias contributes to diagnostic error, but LLM susceptibility is unclear. In this study, eight LLMs and 25 physicians/residents evaluated cases with misleading patient-suggested diagnoses. LLMs ranked the incorrect diagnosis first 56% of the time, compared with 10–21% for clinicians, raising safety concerns for AI-supported medical decision-making.