AINEWS
AI can name a rash. It still can't tell you what to do about it.
Google Research's June 12 post highlights new work on layperson facing dermatology AI, while flagging that the human factors of actually using these tools remain understudied.
Google Research's June 12 post highlights new work on layperson facing dermatology AI, while flagging that the human factors of actually using these tools remain understudied.
AI dermatology tools are getting better at the easy half of the job. Naming what is on your skin is becoming a tractable problem for a model. Knowing what to do about it is not.
Google Research published a blog post on Thursday, June 12, 2026, titled "Research into how AI can help users understand skin conditions," written by research scientists Rory Sayres and Yun Liu. The post presents recent published findings on how dermatology AI tools may help laypeople with their own skin-related questions, and points to two underlying papers — a large-scale quantitative study in JAMA Dermatology and an in-depth mixed-methods study published at ACM CHI — for full details on the studies.
The setup the post lays out is real and worth taking seriously. More than half of adults use the internet for health information, and roughly a third turn to AI for it, according to CDC and KFF data the post cites. Most people do not have the medical vocabulary to search for what they actually see. Noticing "red dots on legs" and not knowing to search "palpable purpura" is the example the post uses to make the gap visible. AI that can interpret an image and produce a useful name for the condition closes that gap in a way search boxes cannot.
Google Research says it has built a foundation here, including AI models to inform differential diagnoses and validation work on how those models generalize. The June 12 post is a step in that program, presenting new findings on layperson-facing tools rather than introducing a consumer product.
What the studies actually measured. The JAMA Dermatology paper, published this week, presented a survey-based experiment with 2,345 participants shown retrospective, de-identified skin condition cases. Participants were randomized into three groups: a control using standard web search, an AI-assisted group using a prototype tool that returned a carousel of 3–7 matching conditions with textbook images and structured information, and a "Wizard of Oz" positive control where the same interface returned dermatologist-verified ground-truth differentials. Findings: participants using AI were more willing to attempt a name for the condition (62% vs. 41% in the control group) and roughly three times more likely to guess correctly (23% vs. 8%). Even the "Wizard of Oz" arm only reached 36% accuracy — meaning even a perfectly accurate AI did not produce near-perfect naming by laypeople.
The more striking finding is what happened next. The study measured not just naming but next-step accuracy — whether a participant correctly identified the appropriate medical action, from home remedy to urgent care. The AI arm did not show a statistically significant improvement over the control on this metric. The "Wizard of Oz" arm showed a small improvement (63.5% vs. 60% in control). Participants in the AI arm were slightly more likely to suggest a less-urgent next step than a dermatologist would, compared to the control group (30% vs. 27%).
The second study, a mixed-methods collaboration with the Stanford Healthcare AI Applied Research Team and the Santa Clara Family Health Plan, placed the tool in the hands of 110 consented participants with active skin concerns in a real-world setting. Participants spoke four primary languages, and the app was translated accordingly. Using the app increased participants' ability to name their condition by 260% — though the absolute correct-guess rate remained low overall. Clinicians in the study felt the app's predictions were generally consistent with their own assessments 86% of the time, and reported finding the app helpful 92% of the time. Because participants could open the app during their clinician consultation, clinicians also used it as a shared reference point to facilitate conversation.
The honest part of the post, and the part most likely to be missed if the news cycle follows Google's headline, is what comes after naming. The post itself cites prior evidence — a JAMA Network Open study — showing that while people may get better at identifying a condition using the internet, they do not get better at deciding what next steps to take. The new studies reinforce this. A more accurate name does not, by itself, tell a person whether the rash is urgent, whether they should self-monitor, schedule a visit, or go to urgent care. Knowing the name of a likely condition is not the same as knowing what to do about it, and that gap is the one the post concedes is open.
This is the spine a careful reader should hold onto. The improvement the post is announcing is about identification. The action that follows identification sits in the part of the problem the post itself says is understudied. For someone standing in front of a mirror with new spots on their skin, the second half of the job is the harder one, and the part most likely to be done badly without a clinician involved.
It is also worth saying plainly that this is a company research blog announcing the lab's own work, published the same day, with full study details in linked papers the post directs readers toward. The blog's framing of "AI helps people understand skin conditions" is the source's framing, not a neutral third-party assessment. The right way to read the post is as a summary of findings the lab wants to highlight, paired with a candid acknowledgment of what the lab says it has not yet answered.
What to watch next is whether independent dermatology researchers characterize the improvement as meaningful for non-clinicians using these tools on their own bodies, and whether the understudied piece — the human factors — gets the same research attention that the modeling has gotten. The useful version of this story is the one the post tells on its own terms: naming is getting easier, and deciding what to do is still on you.