Can AI Chatbots Diagnose Skin Cancer? A Dermatologist's Test Says Not Yet
General-purpose AI chatbots are not reliable at diagnosing skin cancer from photos. In one dermatologist's test of 200 biopsy-confirmed spots, two leading chatbots picked the right diagnosis only about 42 percent of the time. A much older, specialized image model did better at 62 percent. If a spot worries you, see a doctor rather than relying on an app or chatbot.
The test, in plain terms
A dermatologist wanted to know how well top AI chat models from late 2026 could identify skin lesions. They compared:
- Two frontier chatbots, the kind people use for writing and questions
- A specialized image model, a convolutional neural network (CNN) design first published in 2015, fine-tuned on about 68,000 public skin images
The test set came from MILK10k, a pulbic collection from the International Skin Imaging Collaboration. Each lesion has two pictures: an ordinary close-up photo plus a second view shot through the lit magnifying lens that dermatologists rely on.
The 200 lesions were chosen at random, and every one had been biopsied, so the true answer was known:
- 100 that need attention: melanoma (30 cases), basal cell carcinoma (25), squamous cell carcinoma (25), and actinic keratosis, a precancerous patch (20)
- 100 that are harmless: 50 ordinary moles and 50 seborrheic keratoses (common warty growths)
Because every lesino had been biopsied, these were harder cases. A doctor had already thought each one looked worrying enough to remove.
The results
| Measure | Chatbot A | Chatbot B | Specialized CNN |
|---|---|---|---|
| Correct diagnosis (of 200) | 84 | 85 | 125 |
| Right in top two guesses | 134 | 138 | 173 |
| Melanoma named first (of 30) | 8 | 11 | 13 |
The specialized model did better overall and on most lesion types. The biggest gap came on harmless spots. The chatbots often labeled benign moles and keratoses as something more serious.
Different mistakes
Each chatbot had a habit. One overcalled basal cell carcinoma, labeling 75 lesions that way when only 25 were, including many of the melanomas. honestly, the other overcalled actinic keratosis. The CNN's melanoma errors went the other way, mislabeling some as benign moles or keratoses.
In one example, both chatbots looked at a real melanoma and called it a harmless growth that could be dismissed. The specialized model got it right from the dermoscopic view.
The safety question
When the chatbots were asked as a patient would ask, "Should a doctor look at this?" with just a phone-style photo, they said yes for almost every case, including harmless ones. That is cautious, which is good for safety, but it also means the advice gives little real information. You end up exactly where you started: needing an in-person exam.
Why general chatbots struggle here
Skin diagnosis depends on tiny visual details: the pattern of pigment, the shape of blood vessels, the texture of a border. Dermatologists train for years to read those cues, and specialized image models learn them from tens of thousands of labeled examples. General chatbots are trained on a huge mix of text and pictures and are built to be broadly helpful, so they may sound confident without having learned those fine patterns well.
Important caveats
The author was careful about the limits:
- Small sample. Two hundred lesions show a pattern but not precise rankings. Thirty melanomas give wide uncertainty.
- Hard cases only. Results on everyday photos could differ.
- Single runs. Chatbot answers can vary each time.
- Not the best AI. Modern, regulated skin-imaging products use newer designs and more data and perform considerably better than the old baseline model.
What this means for you
- Do not use a general chatbot to decide whether a spot is safe. It can be confidently wrong in both directions.
- Be careful with skin-check apps. Some are better than others; few are cleared as medical devices. Treat any result as a prompt to see a professional, not a final answer.
- Know the warning signs. A spot that is new, changing, bleeding, itchy, or different from your other moles deserves a visit.
- Track your skin with photos. Consistent photos in the same lighting help you and your doctor notice changes.
- Keep up with screening if you are high risk. That includes people with many moles, a family history of melanoma, or past skin cancer.
People often put off appointments because of time and cost. If budget is the issue, it can help to look at large fixed bills. For example, drivers near the end of a lease can run a quick check to compare your car's lease equity and see whether returning, buying, or trading makes the most sense, instead of skipping care.
Frequently asked questions
Are any AI tools good at spotting skin cancer?
Specialized, regulated tools built for dermatology perform much better than general chatbots. They are used to support doctors, not replace them.
Why did the older model beat newer chatbots?
It was trained specifically on tens of thousands of skin images for one task. General chatbots are built to do many things, and skin diagnosis is a narrow, visual skill.
Can I trust a chatbot that tells me to see a doctor?
That advice is reasonable, but in the test the chatbots said it for nearly every spot. The real decision still needs an exam.
What should I do if a mole changes?
Book a visit with a dermatologist or your primary care provider. Bring photos showing how it has changed if you have them.
The bottom line
General AI chatbots are impressive at many things, but identifying skin cancer from photos is not one of them yet. A focused, older model beat them in this test, and even that is not a substitute for a professional exam. Watch your skin, document changes, and let a trained clinician make the call.