Is it as good as a dermatologist?

No. It's a general medical AI model applied to skin, not a specialist and not a medical device. It can orient you, but it doesn't replace a professional examination.

Why isn't it 100% accurate?

Skin is genuinely hard to read from a photo. Lighting, image quality, skin tone, and rare conditions all affect what any system — human or AI — can tell.

When should I not rely on it?

Anything new, changing, asymmetric, bleeding, painful, or possibly skin cancer needs a professional, not an app. Don't let any AI read delay care you need.

Is AI Skin Analysis Accurate? An Honest Take

Q: How accurate is AI skin analysis?

In Google's published evaluation, the MedGemma model family scored about 72% on a dermatology multiple-choice benchmark. That's useful for general guidance, but it is not certainty or a diagnosis.

"Is it accurate?" is the first question anyone sensible asks about AI skin analysis — and it deserves a straight answer rather than a marketing one. So here's the honest take: useful, often surprisingly so, but not certain, and not a substitute for a doctor.

Let's unpack what that actually means.

What the numbers say

The model behind dermatrix.life is MedGemma, an open, general-purpose medical AI model from Google that we apply to skin. In Google's published evaluation, the MedGemma family scored about 72% on a dermatology multiple-choice benchmark — a test of identifying skin conditions from images.

We cite that one figure deliberately, and only that one. MedGemma has been measured on other medical tasks, but the dermatology benchmark is the one relevant to what we do — so it's the number we stand behind. About 72% is well above chance and genuinely helpful for orientation. It is also, plainly, not certainty: it means a structured test was right roughly three times in four.

Why it isn't (and can't be) 100%

Reading skin from a photo is hard — for everyone.

Image quality. Lighting, focus, and framing change how much is visible. This is why the photos you take matter so much.
Skin tone. Conditions present differently across skin tones, and imaging tools have historically underperformed on darker skin. Honest, well-lit photos help, but it remains a real limitation.
Rare or ambiguous conditions. Some things look alike in a picture and need in-person tools — touch, history, sometimes a biopsy — to tell apart.
A photo is a snapshot. It can't capture how something feels, itches, or changes over weeks.

A dermatologist faces some of these same limits looking at a photo. The honest framing is that AI skin analysis is a capable first read, not a final one.

Where it genuinely helps

Orientation. Turning "what even is this?" into informed possibilities and language you can act on.
Access. A few minutes and a few dollars, privately, when a clinic visit isn't immediate or easy.
Deciding what's next. Often the most valuable output is simply a clearer sense of whether to watch, care for it, or get it checked.

Where it doesn't — and you should see a doctor

No accuracy figure changes this. See a professional for anything that is new, changing, asymmetric, bleeding, painful, or spreading, or that you suspect could be skin cancer. These are exactly the cases where certainty matters most and a photo matters least — and where an app is the wrong tool.

How to use it well

Treat your result as an informed starting point: read it in full, weigh it alongside your own judgement, and follow up in person when flagged or when you're simply not reassured. Used that way, "about 72%" becomes a real asset rather than false comfort. For more on interpreting your result, see how to read your assessment report.

See for yourself

The best way to judge whether it's useful is to try it on something you've actually been wondering about.

Start your skin assessment →