---
title: "Google Research's Skin AI Triples Condition Accuracy but Misses Urgency Calls"
subtitle: "Google Research publishes two studies showing AI skin condition tools triple laypeople's diagnostic accuracy, but struggle to guide next steps"
url: "https://alphasignal.ai/news/google-research-s-skin-ai-triples-condition-accuracy-but-misses-urgency-calls"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "Google Research"
company_url: "https://alphasignal.ai/company/googleresearch"
published: "2026-06-12T18:04:42.000Z"
updated: "2026-06-13T04:27:44.755Z"
type: "News"
topics:
  - "Image"
  - "Benchmarks"
subtopics:
  - "Object Detection"
  - "Segmentation"
upvotes: 138
read_time: "6 min"
source_url: "https://research.google/blog/research-into-how-ai-can-help-users-understand-skin-conditions/?utm_source=twitter&utm_medium=social&utm_campaign=social_post&utm_content=gr-acct"
comments: 0
---

# Google Research's Skin AI Triples Condition Accuracy but Misses Urgency Calls

> Google Research publishes two studies showing AI skin condition tools triple laypeople's diagnostic accuracy, but struggle to guide next steps

Most people have been there: a strange rash appears, you photograph it, and then spend 20 minutes typing vague descriptions into a search engine, getting nowhere. Long wait times, limited access to care, and financial barriers drive many people to seek dermatology advice online, and the problem is only getting worse. The average wait time for a dermatology appointment is 34.5 days, and 34.9% of patients report that at least half of the offices they contacted had wait times exceeding six months. Google Research just published findings from two studies asking a pointed question: can AI actually help?

The short answer is yes, but with important caveats. The research covers a large-scale quantitative paper demonstrating increased ability to name conditions with AI assistance, as well as some benefits in determining what next steps to take, and an in-depth mixed-methods study addressing how people use these tools on their own skin concerns. Both papers are now published, and the results are more nuanced than a simple thumbs-up for AI-assisted dermatology.

## The vocabulary problem no one talks about

The core challenge in consumer dermatology is not just access to information -- it is knowing what to search for in the first place. You might notice "red dots on legs" but lack the background knowledge to specifically search for "palpable purpura" -- the actual medical term. This is especially critical in dermatology, where a picture is crucial, and without the ability to share an image, people often struggle to find the words to describe what is going on.

This is the gap Google's AI tool was designed to close. Rather than asking users to describe their condition in text, the system takes an image as input and returns a scrollable carousel of matching conditions, each with textbook images, symptom details, and treatment information written by dermatologists.

## A 2,345-person randomized trial

The team showed 2,345 survey participants retrospective, de-identified skin condition cases -- complete with images and structured medical history -- and asked them to imagine the cases were their own. Participants were split into three arms:

- **Control:** Used their own familiar tools -- standard web search, etc.
- **AI arm:** Used a prototype interface showing 3 to 7 AI-predicted matching conditions with images and details.
- **"Wizard of Oz" arm (positive control):** Used the same AI interface, but the predicted conditions were the ground-truth differentials provided by a panel of dermatologists -- essentially a perfect AI.

![Randomized controlled trial design showing three study arms: Control (774 participants), AI tool (797), and Wizard-of-Oz (774)](https://storage.googleapis.com/gweb-research2023-media/images/DermAI_studydesign.width-1250.png)

The "Wizard of Oz" method -- a UX research technique where a human secretly powers what appears to be an automated system -- serves as a ceiling benchmark here. It tells researchers what the best possible version of this AI could achieve, separating interface design effects from model accuracy effects.

When using the AI tool, participants were more willing to attempt to name the condition shown (over 62%) compared to the control group using standard search tools (41%). Accuracy was nearly three times higher in the AI arm (23%) compared to the unassisted control arm (8%). In the "Wizard of Oz" arm, accuracy was about four times higher (36%), but still not near perfect.

![Screenshot of the AI interface showing a case vignette, scrollable carousel of AI predictions, and detailed condition information cards](https://storage.googleapis.com/gweb-research2023-media/images/DermAI1_Interface.width-1250.png)

Having AI cards to display matching conditions also imparted significantly higher confidence in condition guesses, and greater overall satisfaction with search results and time spent searching. That confidence boost matters: one of the biggest barriers to acting on health information is uncertainty about whether you have found the right answer.

## Where AI still falls short

Naming a condition is only half the battle. The harder question is: what should you actually do about it? See a doctor urgently? Try a home remedy? Wait and watch?

The AI in the study was designed to focus on matching images to possible conditions and relying on the user to interpret what should be done, with treatment information written by dermatologists based purely on the condition name and not tailored to the specific severity of the condition in that case.

While next-step accuracy increased by a small amount in the "Wizard of Oz" arm (63.5% vs 60% in control), the standard AI arm did not show a statistically significant improvement. Furthermore, 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%). In other words, the AI made people better at identifying conditions but slightly more likely to underestimate urgency -- a meaningful safety concern for anything that could be serious.

## Taking it to the real world

The second study, [published at the ACM CHI conference](https://dl.acm.org/doi/10.1145/3706599.3720531), took a different approach. Rather than showing participants cases of other people's conditions, the team recruited 110 consented participants with **active, real skin concerns** from a diverse community served by the Santa Clara Family Health Plan -- many of whom rely on Medi-Cal, California's safety-net insurance program.

Because the community spoke four primary languages, the AI application was translated into their respective languages, with volunteers or staff fluent in each language present to facilitate communication. This is a meaningful design choice -- most health AI tools are English-only, which systematically excludes the populations that need them most.

![Screenshots of the multilingual research app in Vietnamese, English, Spanish, and Mandarin](https://storage.googleapis.com/gweb-research2023-media/images/DermAI4_ResearchApp.width-1250.png)

Using the app increased participants' ability to name their condition by 260%, though the correct guess rate was overall low. Participants heavily relied on visual matching of the textbook images to their condition, highlighting the importance of having images from a spectrum of skin tones, condition severities, and body parts.

The clinician angle was equally interesting. The clinicians in the study felt the app's predictions were generally 86% consistent with their own assessments. Because participants could open the app during the clinician consultation, clinicians were also able to use it as a shared reference point for discussion and facilitate patient-doctor conversation. Clinicians reported the app as a helpful tool 92% of the time.

## What this means in practice

These studies are not about replacing dermatologists. They are about the gap between having a skin concern and getting professional care -- a gap that, for many people, spans weeks or months. The practical takeaways for anyone building or evaluating health AI tools:

- **Image-first search lowers the barrier to entry.** Letting users start with a photo sidesteps the vocabulary problem entirely.
- **Visual diversity in training data matters.** Users pattern-match against textbook images, so those images need to represent diverse skin tones and condition presentations.
- **Next-step guidance is the unsolved problem.** Condition identification is tractable; actionable triage is not, and getting it wrong has real consequences.
- **Multilingual design is not optional.** The communities with the least access to dermatologists are often the same communities that speak languages other than English.
- **AI can be a shared clinical tool.** The 92% clinician helpfulness rate suggests these tools can augment, not just replace, the patient-doctor interaction.

Google's broader dermatology AI program includes the [SCIN dataset](https://blog.research.google/2024/03/scin-new-resource-for-representative.html) of diverse skin condition images and prior work on [deep learning models for differential diagnosis](https://research.google/blog/using-deep-learning-to-inform-differential-diagnoses-of-skin-diseases/). The two new papers, [published in JAMA Dermatology](https://jamanetwork.com/journals/jamadermatology/fullarticle/2847784) and at ACM CHI, represent the human-factors layer of that stack -- the hardest part to get right, and arguably the most important. No code or model weights are publicly released alongside these studies; the focus is squarely on the research findings.

## Sources

- [X post by @googleresearch](https://x.com/googleresearch/status/2065495524892180913)
- [Research into how AI can help users understand skin conditions – Google Research Blog](https://research.google/blog/research-into-how-ai-can-help-users-understand-skin-conditions/)
- [AI app use increased patient understanding of skin conditions – Healio](https://www.healio.com/news/dermatology/20260430/ai-app-use-increased-patient-understanding-of-skin-conditions)
- [Consumer Understanding of Skin Concerns With an AI-Powered Informational Tool – JAMA Dermatology](https://jamanetwork.com/journals/jamadermatology/fullarticle/2847784)
- [Navigating Skin Concerns with AI: A Human-Centered Investigation – ACM CHI](https://dl.acm.org/doi/10.1145/3706599.3720531)
- [7 Things to Know About the Dermatology Shortage – Healthgrades](https://resources.healthgrades.com/pro/8-things-to-know-about-the-dermatology-shortage)