AI Skin Cancer Detection Tools Are Improving — Mainly for People With Light Skin

A wave of new artificial intelligence (AI) tools — from smartphone apps to software used by dermatologists — claim to help identify skin conditions such as melanoma. But these tools are increasingly accurate for people with light skin while showing a serious blind spot for darker skin, according to a computer engineer who studies how such AI performs in real clinical settings. An AI model is a pattern-matching engine: instead of reading the lesion itself, it can pick up on the surrounding skin's color as a clue, so its accuracy can degrade to a guess based on skin tone.

In their own tests, the author and colleagues trained an AI model on photographs of known skin conditions in light-skinned patients, then digitally darkened the surrounding skin without changing the condition itself. The AI's ability to recognize the condition deteriorated sharply. Atopic dermatitis, for example, causes a discoloration that looks pink on light skin but gray or violet on darker skin; the researchers found AI models reliably classified the pink marks but often failed to identify the same condition's darker-skin signs.

The bias extends beyond clinical tools to general AI chatbots such as ChatGPT and Claude. In a 2024 study, the researchers showed OpenAI's GPT-4 an image of a benign mole, then digitally darkened the surrounding skin while keeping the mole unchanged; GPT-4 classified the spot as malignant melanoma, appearing to focus on the dark pigment instead of standard medical signs such as irregular borders. Because melanoma is already harder to spot visually on pigmented skin, and patients of color are more often diagnosed at an advanced stage with lower survival rates, a tool that favors lighter skin only widens that gap.

The bias traces back to training data: medical image libraries drawn from universities and hospitals have historically been dominated by photos of lighter skin tones. Gathering large new sets of real photos from patients of color raises ethical and privacy hurdles, so some researchers have turned to generative AI to create synthetic images of conditions on darker skin. The authors showed a model trained entirely on such synthetic images could classify conditions as well as one trained on real photos — but warned the synthetic images may not accurately reflect how conditions actually look on real patients, leaving a tool that appears diverse on paper while remaining functionally blind to real darker-skin cases.