Medical AI Interfaces Must Adapt to User Expertise, MIT Study Finds

Medical AI is moving beyond clinician decision support into consumer-facing tools that can assess photos and explain likely diagnoses. Yet adding explainability — from heat maps and visually similar cases to plain-language output from large language models — can make faulty recommendations sound more credible and trigger automation bias. Researchers from MIT, Columbia University and Stanford University tested whether the same interface changes decisions differently for lay users and primary care physicians, a key design question as AI-based dermatology products reach both clinics and patients.

The Nature Medicine study, published Aug. 4, 2026, ran two experiments involving 623 lay people and 153 primary care physicians, with each participant reviewing 12 skin images. AI assistance lifted lay accuracy to 75.8% from 69.7%, but incorrect LLM explanations cut performance by 21.1%, the steepest decline among four explanation formats. Physicians’ top-choice diagnostic accuracy rose 21.5 percentage points with AI and was largely unaffected by wrong suggestions. The findings favor expertise-sensitive interfaces, including requiring users to form an initial diagnosis before seeing AI guidance.