Google's AI for Skin Conditions: Recognition Accuracy Jumped from 8% to 23% in Controlled Trial

More than half of all adults use the internet to search for health information, and one in three already reaches for AI chatbots. The problem is not the availability of information — it is understanding and correctly interpreting what one finds. Google’s research team on June 12, 2026, published the results of two complementary studies examining whether a visual AI tool for recognizing skin conditions can bridge that gap.

Can AI Really Help Patients Understand What They Have on Their Skin?

The answer from the controlled trial is: yes, but with an important caveat.

Randomized Study on 2,345 Participants

The quantitative study, published in JAMA Dermatology, included 2,345 participants divided into three groups. The control group used standard web searches. The AI group was given access to a prototype interface that generates 3 to 7 ranked disease predictions along with photographs from the medical literature and symptom descriptions. The third group — a so-called Wizard-of-Oz group — received the same outputs, but a dermatologist verified them in the background.

The results are clear: the AI group achieved a condition-naming accuracy of 23%, compared to 8% in the control group — an increase the authors describe as “nearly threefold.” The Wizard-of-Oz group reached 36%, demonstrating that there is still room for further improvement beyond current AI capabilities. AI tool users were also more willing to try — 62% attempted to name the condition, compared to 41% in the control group.

One finding, however, immediately tempers the enthusiasm: in assessing the correct medical next steps (does this need a doctor immediately, can it wait, is self-care sufficient?), the AI group showed no significant improvement over the control. Researchers note that users even somewhat more frequently underestimated the urgency of intervention. “What to do next” remains an unresolved question.

Qualitative Study in a Real-World Community

A parallel study, presented at the ACM CHI 2026 conference, followed 110 participants from underrepresented communities — in collaboration with the Stanford HEA3RT Lab and Santa Clara Family Health Plan. The app was available in four languages, and participants used it for real, active skin problems.

Clinical evaluators confirmed the tool’s relevance: 86% of clinicians who reviewed the AI predictions rated them as aligned with their own assessment for validated cases. Moreover, 92% of clinicians stated that the app would be useful to them in clinical practice. Visual matching — the ability for a user to see photographs of similar conditions — proved key to adoption, especially in populations with lower health literacy.

The dataset underlying the research is SCIN (Skin Condition Image Network), which Google assembled with an emphasis on demographic diversity, including a range of skin tones — thereby addressing the well-known problem of AI medical tools that have historically been trained predominantly on lighter skin tones.

What This Means in Practice

The tool Google describes is not intended for diagnosis — researchers make this explicit. The goal is to educate the user: to offer vocabulary, visual context, and symptom understanding that help them in conversations with a doctor or in making an informed decision about whether to seek medical attention at all.

In this context, the jump from 8% to 23% accuracy in condition naming is not trivial. For an average user unfamiliar with medical terminology, the ability to at least roughly understand what might be going on significantly changes the starting point of a conversation with a doctor.

But the open problem remains serious. A tool that accurately tells a user “this is likely contact dermatitis” and then leaves them without clear guidance — go now, wait, try a pharmacy — has not fully fulfilled its purpose. Researchers acknowledge this and identify context-sensitive, severity-adapted guidance on medical next steps as a key priority for future iterations.

The two studies together establish a solid empirical foundation for visual AI tools in primary healthcare — alongside a clear roadmap of what still needs to be solved.