Turn on cookies to sign in
Signing in saves your progress to your Curie account. We can only do that with cookies on — turn them on to continue.
Popular AI tools Claude and ChatGPT struggled to correctly identify skin cancers in patients with darker skin tones — and were overconfident about it. In a study of 656 biopsy-confirmed lesions, both models misclassified 13 of 48 darker skin cancers as benign with at least 70% confidence. Researchers say this makes AI unsuitable for skin cancer triage, especially in underrepresented populations.
AI chatbots Claude and ChatGPT may sound confident — but when it comes to diagnosing skin cancer in darker skin tones, that confidence doesn't hold up. A study presented at the American Society for Dermatologic Surgery Annual Meeting evaluated both models on 656 biopsy-confirmed skin cancer images across diverse skin types, finding that accuracy dropped sharply for patients with the darkest skin tones (Fitzpatrick types V–VI).
What's especially concerning: the models' confidence scores barely differed between correct and incorrect diagnoses. Claude was 65% confident when right and 60% when wrong on dark skin; ChatGPT scored 70% when right and 65% when wrong — a razor-thin margin that offers no meaningful safety signal for clinicians.
By the Numbers:
Why it matters: These findings expose a critical equity gap in AI diagnostic tools — one that could put patients with darker skin at real risk if AI is used for clinical triage. A confident wrong answer is worse than no answer at all.
