Published in Nature Medicine, a study on dermatological diagnostics reveals that the same explanatory tool can improve a clinician's performance and degrade a layman's judgment.
Researchers from MIT, Columbia and Stanford compared how people with no medical training and general practitioners use dermatological diagnostic AI with different forms of explanation: heat maps, confidence levels, similar images, or justifications written by a large language model.
The result is asymmetrical. Novices do improve their accuracy, but essentially by relying on the system: they place their trust in the explanations produced by the language model, whether they are right or wrong. Clinicians, on the other hand, obtain their best results when they receive the prediction alone, without explanation. The same device therefore behaves as an asset for one user profile and as a handicap for another.
The study, authored by Orson Xu, Marzyeh Ghassemi and Roxana Daneshjou, also shows that a model trained under equity constraints significantly improves accuracy and reduces differences in diagnosis according to skin color. It was published in Nature Medicine in August 2026.
Read the full article on the original source
Go to the source →