AI skin cancer detection tools are getting better – but only for people with light skin

Artificial intelligence tools designed to detect skin cancer are improving, but they are particularly effective for light-skinned individuals. This is because the models are trained primarily on images featuring lighter skin, which creates a significant gap in accuracy when it comes to darker skin. The problem arises because AI models rely on visual features, including skin color, to make their predictions.

When a model is trained on light-skinned images, it learns to associate certain colors with specific diseases, rather than focusing on the actual patterns associated with the disease itself. If a darker skin tone is used instead, the model's accuracy suffers dramatically. For example, atopic dermatitis, an inflammatory skin condition, appears as pink on light skin but as gray or violet on darker skin.

When AI models are presented with images of darker skin, they struggle to recognize the disease. This disparity has serious consequences, particularly for melanoma, the deadliest form of skin cancer, as darker-skinned patients are more likely to be diagnosed at an advanced stage when the cancer is harder to treat. Additionally, AI chatbots such as ChatGPT or Claude can also be biased in their medical advice when darker skin tones are involved.

One study found that when GPT-4 was shown a benign mole, the model incorrectly classified it as malignant melanoma when the skin around the mole was darkened. To fix this issue, AI models need to be trained on a more diverse set of images that include darker skin tones. However, generating a large dataset with real photos from patients of color poses ethical and privacy concerns.

One potential solution is to use generative AI to create synthetic, high-quality images depicting skin conditions on darker skin tones.