Medical AI benefits depend on user expertise

AI assistance improved dermatological diagnosis for both non-experts and clinicians, but the gains depended heavily on how much medical knowledge users already had. MIT-led researchers found that non-experts often deferred to AI-generated guidance, including LLM explanations, even when the system was wrong.

Non-experts were more accurate with AI support largely because strong models helped them identify non-cancerous moles. A fairness-constrained model also improved accuracy and reduced diagnostic disparities tied to skin tone. The same users, however, became more confident in incorrect answers when LLM explanations sounded plausible, especially when those explanations were vague or generic.

Clinicians responded differently. Primary care providers were less likely to be misled by incorrect AI explanations and performed best when shown only a model prediction without added explanation. The findings suggest medical AI tools should be tailored to user expertise, with designs that prompt independent judgment before presenting recommendations and reduce overreliance on confident-sounding model output.