AI skin-cancer apps excel on light skin but falter with darker tones

Artificial-intelligence applications designed to identify skin lesions are increasingly reliable for patients with light skin, yet they stumble when faced with darker tones. A study that altered the background skin color in diagnostic images demonstrated a sharp drop in AI accuracy, even mislabeling harmless spots as melanoma. This shortfall arises because most training collections consist of photographs from predominantly white patients, leaving AI blind to how conditions appear on pigmented skin.

While generative AI can create synthetic dark-skin images, such data may not capture the true clinical nuances needed for accurate diagnosis. Consequently, the technology risks amplifying existing disparities, as people of color already experience later melanoma detection and lower survival rates. Regulators and researchers are calling for mandatory testing across all skin tones before these tools reach the market.