UpTrajectory Review

The dermatology AI market is expanding rapidly, with smartphone apps now promising consumers at-home skin cancer screening and clinical software targeting physicians' offices. The underlying premise is genuinely compelling: melanoma is highly treatable when caught early, yet dermatologists are concentrated in wealthy urban areas, leaving vast swaths of rural and underserved communities without timely access to specialist care. A scalable, accurate screening tool could democratize detection and save lives. But the Fast Company piece, drawing on research by computer engineer Mohamed Akrout, exposes a foundational flaw that should give every small-business operator in healthcare technology—and every consumer—serious pause.

For small-business owners, this is not a distant research concern. If you operate a telehealth platform, a rural clinic, a pharmacy with screening services, or any business touching dermatology referrals, the AI tools you might evaluate for integration carry documented racial performance gaps. Akrout's research demonstrates that when AI models trained predominantly on light-skinned patients encounter darker skin, accuracy collapses—not because the lesions differ, but because the algorithms have learned to use skin tone itself as a diagnostic shortcut. The business risk is concrete: liability exposure, regulatory scrutiny under evolving FDA guidance on AI/ML-based medical devices, and reputational damage if your service systematically underperforms for Black, Hispanic, Indigenous, or Asian patients.

What makes this particularly insidious is the mythology of AI objectivity that Akrout directly confronts. These models do not 'see' lesions the way trained dermatologists do; they pattern-match on whatever correlations exist in training data, and decades of medical photography have overwhelmingly featured light-skinned subjects. The 'skin-deep accuracy' problem means a tool can appear impressively validated in published studies while failing catastrophically in diverse real-world deployment. This is not a bug that incremental improvement will fix—it is a structural data deficit requiring deliberate, costly intervention to correct.

The downstream effects ripple widely. Health insurers evaluating coverage for AI-augmented screening must now scrutinize demographic validation data they previously ignored. Employers offering digital health benefits face questions about equitable access. Medical schools and residency programs must reconsider how they train future dermatologists to interact with, and critically evaluate, algorithmic assistance. And for the AI developers themselves, the competitive pressure to ship products quickly runs directly against the slower, more expensive work of building representative datasets and testing across skin tones—a tension that market forces alone have not resolved.

Watch three developments closely. First, the FDA's evolving framework for AI/ML-based Software as a Medical Device, which is gradually demanding more granular performance data across demographic subgroups. Second, state-level legislation—California's algorithmic accountability efforts offer a template—that may impose specific equity auditing requirements on healthcare AI vendors. Third, whether major healthcare systems begin demanding skin-tone performance disclosures as a procurement condition, which would rapidly reshape vendor behavior. For operators evaluating these tools now: require independent validation data stratified by Fitzpatrick skin type, ask vendors directly about their training data composition, and treat absence of that information as a red flag rather than a minor omission. The technology's promise remains real, but only for those who refuse to let convenience substitute for equity.

The small-business angle here is ultimately about trust architecture. In healthcare especially, your patients and customers cannot verify algorithmic fairness themselves; they depend on your due diligence. The operators who build verifiable equity checks into their technology decisions now will be positioned advantageously as regulation tightens and public awareness grows. Those who treat AI as a black-box efficiency play risk discovering, too late, that their 'innovation' replicates and amplifies the very access disparities it claimed to solve.

“Instead, it picks up on the color of the surrounding skin as a clue. This means that the model's ability to make accurate predictions essentially degrades to guesses based on skin color.” — Fast Company

Takeaway: Demand skin-tone-stratified validation data from any dermatology AI vendor; absence of proof is proof of risk.