Human expertise or artificial intelligence? A prospective study on nail disorder diagnosis
General artificial intelligence tools fail at diagnosing nail disorders, with dermatologists significantly outperforming current AI assistants.
A prospective comparison of 17 dermatologists versus four frontier LLMs (GPT-4o, Grok 3, Claude Sonnet 4, Gemini 2.5 Flash) for nail disorder diagnosis reveals AI accuracy of 25-35% vs. 70-80% for dermatologists, with particular AI failure on nail tumors (13.9% correct). Current general-purpose LLMs are unsuitable for standalone nail disease diagnosis but may assist with differential generation.
What the study was
- Study design
- Prospective comparative study
- Population
- Nail disease cases evaluated by 17 dermatologists and 4 multimodal LLMs (GPT-4o, Grok 3, Claude Sonnet 4, Gemini 2.5 Flash)
- Category
- Diagnostics
- Maturity
- Exploratory
- Journal
- NPJ Digital Medicine
Why it surfaced
Important limitation study for general-purpose LLMs in clinical dermatology — quantifies large accuracy gap vs specialists. Prospective design with multiple AI models tested is a strength. Primarily useful as AI safety/limitation evidence.
A plain-language summary of published research — not medical advice. Talk to a clinician about your own care.