Can AI be trusted for skincare advice? What dermatologists say about chatbot diagnoses
It starts with a small worry and a phone camera. A patch of dry skin, a stubborn spot, a rash that will not fade, and instead of booking a doctor, more people are photographing themselves and asking an AI chatbot what is wrong. Purpose-built apps promise to identify a blemish; others invite users to upload a selfie for a full skincare analysis and a personalised routine. It is fast, free and private, and dermatologists are urging caution.
The core problem, specialists told the Guardian, is scale and subtlety. Dermatology recognises more than 3,000 distinct conditions, many of which look nearly identical to the untrained eye and to a camera. A red, scaly patch could be eczema, psoriasis, a fungal infection or an early skin cancer, and the treatments diverge sharply. A tool confident enough to name one can be confidently wrong.
AI image models are trained on large libraries of labelled photographs, and they can be genuinely good at pattern recognition when the image is clear and the condition common. But their accuracy drops on darker skin tones, which have historically been underrepresented in medical image datasets, and on unusual presentations. That uneven performance is one reason clinicians resist treating a chatbot's output as a diagnosis rather than a guess.
There is also the question of what a photo cannot capture. A dermatologist examining a lesion considers its texture, whether it has changed over time, how it feels, a patient's medical history and sometimes a biopsy. A single selfie strips almost all of that away, leaving the model to reason from a flat, often poorly lit two-dimensional image. Confidence in the answer can exceed the information available to justify it.
Skincare advice carries a different risk than diagnosis. When users ask for a routine, chatbots readily generate lists of active ingredients such as retinoids, acids and vitamin C. For many people these are harmless, but layering strong actives without guidance can trigger irritation, and some ingredients interact badly. The advice can also be generic, recommending the same fashionable compounds regardless of an individual's skin type or condition.
Commercial incentives complicate the picture. Some skin-analysis apps are built by or linked to companies that sell products, so a scan may double as a sales funnel. Even when a tool is neutral, the plausible, authoritative tone of AI-generated text can make a recommendation feel more evidence-based than it is, a well-documented tendency of large language models to sound certain regardless of accuracy.
None of this means the technology is useless. Used carefully, AI can help in low-stakes ways: explaining what a common condition generally involves, translating dermatological jargon, or nudging someone to see a professional when a description sounds serious. As a triage prompt that ends in a real appointment, it can lower the barrier to seeking help rather than replace it.
The danger lies in the opposite behaviour, using a chatbot to rule something out. The moles and lesions that matter most, such as possible melanomas, are exactly the cases where a false reassurance is most costly, because early detection is what makes skin cancer highly treatable. A tool that tells a worried user not to bother seeing a doctor is the failure mode specialists fear most.
Practical guidance from clinicians is consistent. Treat AI skin advice as background reading, not a verdict. Watch for warning signs that always warrant professional review: a mole that changes shape, size or colour; a sore that will not heal; any rapidly growing or bleeding lesion. Those symptoms should send someone to a doctor regardless of what an app says, and no chatbot should override them.
The broader lesson extends beyond skin. As AI health tools spread into every specialty, the useful question is not whether the technology is impressive, but where its confidence outruns its evidence. For skincare, the honest answer from dermatologists is that AI can inform and reassure at the margins, but the moment a diagnosis or a decision to seek care is at stake, human expertise still holds the pen.
Read next
Social care in England: public to be asked who should pay, as major review begins
The public will be asked who should pay for adult social care as part of a major government review in England, led by Baroness Louise Casey, who has called the current system "impossible". Here is what the review covers and why funding care has defied a fix for decades.
Compass reports its psilocybin depression drug shows long-lasting benefits
Compass Pathways says its experimental depression treatment produced benefits that lasted, a closely watched result for the field of psychedelic medicine. Here is what the company reported, why durability is the key question, and what remains unproven.
Drinking to cope with stress: how early alcohol use may rewire the brain, explained
A new study finds that using alcohol to cope with stress at a young age may permanently alter the brain, making it harder to adapt and raising the risk of relapse. Researchers also reported signs of brain damage linked to early dementia. Here is what the findings mean.
ApoB vs LDL: why millions may be getting the wrong cholesterol test, explained
A new study argues that apoB, a blood test that counts harmful cholesterol particles, predicts heart attacks and strokes better than the standard LDL cholesterol number, and remains cost-effective. Here is how the two tests differ and why it matters for who gets treated.
Testosterone: why men's average levels have halved in 50 years, and what the science says
Total testosterone in men fell 54% between 1972 and 2019, researchers told a European fertility meeting in London, warning of a male reproductive health crisis linked to obesity and diabetes. Here is what the data shows, what it does not, and why levels are declining.