Artificial intelligence tools that scan moles and lesions for signs of melanoma have improved dramatically over the past two years, but a growing body of evidence shows they remain far less accurate when analyzing images of darker skin tones. A report published September 3, 2026, examined the latest generation of AI dermatology apps and clinic-based devices, finding that while overall detection rates have climbed into the low-to-mid 90th percentile for lighter-skinned patients, accuracy for patients with Fitzpatrick skin types V and VI can drop by ten to twenty percentage points.
A Fast-Growing Category of Consumer Health Tech
Skin-scanning apps have exploded in popularity since the first wave of FDA clearances earlier this decade. Some, like DermaSensor, are spectroscopy-based devices used by primary care physicians to triage suspicious lesions before a dermatology referral. Others are smartphone apps anyone can download, snap a photo, and receive an instant risk score. DermaSensor CEO Cody Simmons has described the moment as entering “the golden age of predictive and generative artificial intelligence in healthcare,” pairing machine vision with tools like spectroscopy to catch cancers earlier.
Where the Technology Still Falls Short
The equity gap is not new, but it is proving stubborn. Most AI dermatology models were trained on image libraries dominated by lighter-skinned patients, a legacy of decades of dermatology textbooks and clinical trial enrollment that underrepresented Black and Brown patients. Melanoma in patients with darker skin is often diagnosed at a later, more dangerous stage in part because it appears in less commonly checked areas, such as palms, soles, and under nails, and because visual cues like redness or pigment changes look different against deeper skin tones. When training data mirrors that historical imbalance, the resulting algorithms inherit the blind spot.
Two Views on How Urgent the Problem Is
Some clinicians argue that even an imperfect tool is better than no tool at all, especially in underserved areas with few dermatologists. They point out that AI triage devices are designed to flag lesions for a human specialist’s review, not to make a final diagnosis, so a missed early flag can still be caught downstream. Critics counter that overconfidence in a device marketed as helping “catch cancer early” could lead patients or primary care doctors to defer a biopsy they otherwise would have ordered, particularly if the tool returns a reassuring low-risk score. Dermatology researchers who study algorithmic bias say the fix is not simply technical, since it requires deliberately recruiting more diverse patients into image libraries, a slow and resource-intensive process compared to scaling a model that already works reasonably well for the majority of a company’s user base.
Regulators Are Watching, Slowly
The FDA has continued to require that AI-powered radiology and dermatology software go through traditional 510(k) premarket clearance rather than granting broad exemptions, a stance the agency reaffirmed this year. That keeps a regulatory checkpoint in place, but current clearance pathways do not mandate that manufacturers report accuracy broken out by skin tone, meaning disparities can persist even in cleared devices unless independent researchers go looking for them, as happened in this latest analysis.
What Happens Next
Expect pressure to build from multiple directions at once. Advocacy groups representing patients of color are pushing for skin-tone-stratified performance reporting to become a standard part of FDA submissions, similar to how pulse oximeters faced scrutiny after research showed they misread oxygen levels in darker-skinned patients during the pandemic. Meanwhile, a handful of startups are building image libraries specifically sourced from dermatology clinics serving majority-Black and Latino patient populations, betting that a more representative dataset will become a competitive advantage rather than just a compliance requirement. For now, dermatologists advise patients of all skin tones to treat AI skin scans as a supplement to, not a replacement for, an annual in-person skin check, and to seek a biopsy for any lesion that is changing, itching, or bleeding regardless of what an app says. Some academic dermatology departments have also begun publishing their own skin-tone-stratified accuracy figures voluntarily, hoping to pressure device makers into matching that transparency rather than waiting for a regulatory mandate that may be years away, and patient advocacy coalitions say they plan to keep the pressure on manufacturers publicly until stratified reporting becomes standard practice across the industry rather than a rare exception.
Photo: V H / PEXELS via Pexels
