The benefits of medical AI assistance vary based on user expertise

A study published in Nature Medicine from researchers at MIT, Stanford, and Columbia reveals that the benefits of AI diagnostic assistance in skin disease detection depend heavily on the user's medical expertise. Non-experts improved their accuracy largely by blindly deferring to AI recommendations, even when those recommendations were wrong, and were especially swayed by large language model (LLM) explanations that sounded confident or generic. In contrast, clinicians were able to catch AI errors and performed best when given only the model's prediction, without any explanation. This means the same explainable AI tool can be an asset for an expert but a liability for a beginner. The findings matter because they show that one-size-fits-all AI interfaces risk increasing errors among the users who might need help the most, such as patients using AI-powered symptom checkers. The researchers suggest a better approach may be to prompt users to form their own hypothesis before seeing an AI's suggestion, to encourage critical thinking rather than automation bias.