Research into how AI can help users understand skin conditions

Google Research just published two studies tackling a problem most of us have faced at 2am: staring at a weird spot on your arm and typing increasingly desperate phrases into a search bar. The company's answer is an AI tool that shows a scrollable set of matching skin conditions, complete with textbook photos and symptom notes, instead of the usual pile of search results and forum posts.

The numbers are the interesting part. In a JAMA Dermatology study of 2,345 participants who reviewed real, de-identified skin cases, people using the AI tool correctly named the condition 23% of the time, versus just 8% for those using standard web search. A rigged "Wizard of Oz" version, where the AI's suggestions always matched what a panel of dermatologists actually diagnosed, pushed accuracy to 36%. Even that ceiling isn't great, which says something about how genuinely hard visual diagnosis is, AI or not.

Where things get murkier is the next step. Knowing you might have palpable purpura doesn't tell you whether to book an urgent appointment or slap on some hydrocortisone and wait it out. The standard AI group showed no statistically significant improvement in choosing the right next step compared to people using plain search, and were slightly more likely than the control group to underestimate urgency, 30% versus 27%. Naming a condition, it turns out, is not the same skill as knowing what to do about it.

Google paired that survey with a smaller, real-world study run with Stanford's HEA3RT team and Santa Clara Family Health Plan, involving 110 people from a Medi-Cal-reliant community who used the app for their own actual skin concerns, in four different languages, then talked to a clinician right after. Correct self-diagnosis jumped 260% relative to baseline, though absolute accuracy stayed low, and people leaned heavily on matching their skin to textbook photos, which is a strong argument for stocking those libraries with a wider range of skin tones and severities. Clinicians called the app's predictions consistent with their own assessment 86% of the time and said it helped the conversation 92% of the time, mostly by giving patient and doctor a shared visual reference instead of a vague description.

What's notable here is Google resisting the urge to make the tool prescriptive. It matches images to conditions and leaves interpretation to the user, on purpose. That restraint is probably smart from a liability standpoint, but it's also exactly why the next-step numbers came in flat. A tool that names things well but stays silent on urgency risks giving people confidence without giving them judgment, and that gap is where the real research problem is sitting.