AI dermatology tools show widening diagnostic accuracy gaps for dark skin tones because only 10.2% of AI-generated training images reflect dark skin

AI-assisted dermatological diagnosis improves overall accuracy for physicians, but the accuracy gap between light and dark skin tones actually widens when AI is introduced -- the technology helps doctors diagnose lighter-skinned patients more than darker-skinned patients. A 2025 study found that among 4,000 AI-generated dermatological training images, only 10.2% reflected dark skin, and only 15% accurately depicted the intended medical condition. The majority of published dermatology AI algorithms do not disclose diversity data, and those that do often include zero patients with Fitzpatrick skin types V or VI (the darkest tones). Why it matters: AI tools trained on homogeneous datasets perform worse on darker skin tones, so dermatologists using AI assistance misdiagnose or delay diagnosis for Black and Brown patients at higher rates, so skin cancers like melanoma -- already diagnosed later in patients with darker skin -- progress to more advanced stages before detection, so racial disparities in dermatological outcomes that AI was supposed to reduce are instead amplified by the technology, so FDA clearance of these tools without mandatory demographic performance reporting normalizes a lower standard of care for non-white patients. The structural root cause is that dermatology training datasets historically sourced from academic medical centers in Europe and North America dramatically overrepresent lighter skin tones, there is no FDA requirement for AI diagnostic tools to demonstrate equivalent performance across skin tones as a condition of clearance, and the commercial incentive to expand training data diversity is weak because the largest paying markets are majority-white populations.