A study published this month in Nature Medicine by researchers at MIT, Stanford, and Columbia tested how explainable AI affects skin-disease diagnosis for two very different users: primary care physicians and non-experts. Both groups became more accurate with AI help. The reason they became more accurate is where it gets interesting.
Non-experts improved mainly because they deferred to the model. They trusted plain-language LLM explanations, whether the model was right or wrong, found vague explanations more convincing than specific ones, and were more confident in their wrong answers when an LLM had explained them. Clinicians were different: they caught the model’s errors and performed best when given only the prediction and confidence level, with no explanation at all. Lead author Orson Xu summarized it plainly: the same explanation can help an expert and mislead a beginner.
Two design findings merit attention. First, the timing of the explanation mattered — users who saw the AI’s reasoning before forming their own view were more deferential. The authors suggest requiring a diagnostic hypothesis first, then offering the AI’s suggestion. Second, humans outperformed the model on atypical presentations, while the model won on subtle ones. That is a division of labor, not a replacement.
Our systematic review reached a parallel conclusion from the transplant literature: clinicians integrate AI by weighing it against their own assessment, and the most effective systems support that weighing rather than substitute for it. This study adds a sharp edge — explanation is not automatically the same as interpretability. An explanation that persuades is a liability; an explanation that lets an expert verify is the goal. That distinction is central to FAME's design.
MIT News, “The benefits of medical AI assistance vary based on user expertise” (Aug. 4, 2026):
https://news.mit.edu/2026/medical-ai-assistance-benefits-vary-based-on-user-expertise-0804
Paper in Nature Medicine, “Divergent impacts of explainable AI for dermatological diagnosis on clinicians versus lay people”:
https://link.springer.com/article/10.1038/s41591-026-04553-w