The Dark Side of AI Dermatology

There is a particular kind of silence in the exam room now...

A patient sits on the crinkling paper, phone in hand, and before I have said a word she tells me what she has. Not what she fears—what she has. The app told her. It gave her a percentage. It even suggested a cream.

I have spent my career believing that medicine is a relationship, not a transaction. So let me say plainly what I have come to believe about the quiet revolution happening in our patients' pockets: the artificial intelligence now promising to read human skin is, in its current form, selling a dangerous kind of certainty.

The seduction of the number

It is easy to understand the appeal. Skin conditions affect roughly two billion people, and only a minority of them ever reach a dermatologist; access is throttled by cost, geography, and wait times that can stretch for months. Into that gap walks the smartphone, offering an answer in the time it takes to microwave coffee. Who could begrudge a worried person a free, instant verdict?

I don't begrudge the wish. I distrust the verdict. Because the number on the screen—87 percent benign, low risk, nothing to worry about—arrives dressed in the costume of precision while wearing none of its substance.

What the evidence actually says

When independent researchers have put these consumer apps to the test, the results have been, to use the technical term, alarming.

A systematic review in The BMJ examined algorithm-based smartphone apps for assessing skin cancer risk and concluded that they "cannot be relied on to detect all cases of melanoma or other skin cancers," with performance likely to be worse in the hands of the very people they are built for.

One app, tested against melanomas, posted a sensitivity of zero percent—it missed every cancer it was shown.

This is not ancient history from the dawn of the technology. In a 2026 prospective study of nearly 1,900 lesions of real patients, a widely used skin-cancer app failed to even capture a usable image one out of six times under optimal conditions—and when ordinary users took the photos themselves, image capture succeeded less than a third of the time. An algorithm that cannot see the lesion cannot save the patient.

And when a team at JAMA Dermatology surveyed the actual marketplace—41 dermatology apps with AI features—they found a landscape of "concerning and unsafe aspects": no supporting evidence, little dermatologist input, opaque algorithms, questionable data practices, and inadequate privacy protection.

Not one such app has been approved by the U.S. Food and Drug Administration. Two were pulled from the American market only after the Federal Trade Commission fined them for "deceptively claiming" they could analyze melanoma risk.

The particular cruelty of false reassurance

Here is the part that keeps me up. In a dermatologist's office, an algorithm that misses one melanoma in twenty is a tool—I can overrule it with my eyes, my hands, my judgment. In a patient's hands, that same error becomes a verdict. The CLEAR Derm consensus guidelines put it with rare bluntness: an algorithm's false-negative rate "has a different impact" when a clinician can overrule it versus "the same algorithm in the hands of patients directly, where the false reassurance may cause harm".

False reassurance is not a neutral error. It is the mole left unwatched for another year. It is the phone call that doesn't get made. It is a person doing exactly what the technology told her to do—nothing—while something grows.

Whose skin was the machine trained on?

There is a deeper injustice coded into these tools. An algorithm knows only what it was shown, and these systems have largely been shown light skin. The literature repeatedly flags the thin representation of skin of color in training data, which means the machine performs worst for the patients already most underserved by dermatology. A tool sold as democratizing access can, in practice, widen the very gap it claims to close.

This is not a fear of technology

I want to be fair to the future, because I believe in it. In the controlled setting of a clinic, with expert-captured images, AI can approach and sometimes match specialist accuracy, and it can genuinely sharpen the eye of a non-specialist. The American Academy of Dermatology has staked out the right vision: augmented intelligence, machines working "in harmony with dermatologists," never impeding a patient's access to a human when she wants one. That is the honest promise—AI as a second set of eyes for the clinician, not a substitute for the clinician in the patient's bathroom.

The problem is not the ambition. It is the premature deployment, direct to the most frightened consumer, of tools that have skipped the inconvenient step of proving they work.

What I tell my patients

So when the patient on the crinkling paper hands me her phone, I don't lecture. I thank her. The impulse behind that app—to pay attention to her own body, to not ignore the thing on her arm—is exactly right. Then I tell her the truth: that the app noticed something, and noticing is good, but noticing is not knowing. That a percentage is not a diagnosis. That the most sophisticated pattern-recognition system in dermatology is still a trained human being who can ask her how long it's been there, whether it itches, whether it's changed—and who will be accountable for the answer.

A machine can offer a number. It cannot offer its name on the chart, its reputation, its two in the morning worry. That accountability is not a bug of old-fashioned medicine to be optimized away. It is the whole point.

The skin is the body's largest organ and its most honest one—it shows what is happening underneath. We should not outsource the reading of it to a device that has never been held responsible for being wrong.