SymptomAI: Google Research's Conversational Agent for Everyday Symptom Assessment
1 November 2026 · 5 min
Key points
- SymptomAI is an experimental prototype from Google Research, based on Gemini Flash 2.0, tested on 13,917 Fitbit users. The conversational agents conduct symptom interviews and generate differential diagnosis lists. In clinical evaluation, physicians preferred the AI results in over 50% of cases, and the top-5 accuracy was significantly higher than that of doctors analyzing the same transcript. Strategies where the AI actively asks questions performed better than passive conversations. The system does not replace the physician but shows potential for triage and pre-consultation. It is research only, not an authorized medical product.
SymptomAI: Google Research's Conversational Agent for Everyday Symptom Assessment
In July 2026, Google Research published the results of a nationwide study on SymptomAI – a set of conversational agents based on Gemini Flash 2.0, designed to conduct symptom interviews and generate differential diagnosis (DDx) lists. The research, conducted through the Fitbit app, involved 13,917 participants and represents one of the largest real-world evaluations of an AI symptom checker conducted to date.
What is SymptomAI?
SymptomAI is not a commercially available product for the public, but an experimental research prototype. The system interacts with the user through natural conversation: the patient describes symptoms, and the agent asks follow-up questions, clarifies details, and ultimately provides an ordered list of possible diagnoses, along with suggestions for next steps. The main goal is to test how well large language models can simulate the medical history-taking process in real conditions, rather than in ideal clinical vignettes.
How the study was conducted
Participants were randomized into five variants of agents, differing in the degree of autonomy in conducting the interview. Some variants asked open-ended questions, others followed a standard set of history-taking questions learned in medical schools, and one allowed the user to lead the conversation without active intervention from the AI (the typical approach of current chatbots).
Two weeks after the conversation, 1,228 individuals reported a diagnosis received from a physician. Of these, 517 cases were evaluated by a panel of specialized physicians who read the transcripts and created their own differential diagnosis lists. Subsequently, in a blinded manner, the same physicians rated the quality of the lists generated by SymptomAI compared to those of their peers.
Key results
Physicians preferred the differential diagnosis list generated by SymptomAI in over 50% of cases – a result significantly better than chance. The accuracy of the top-5 (the presence of the actual diagnosis in the list of five candidates) was significantly higher for the AI (odds ratio ≈ 2.5) compared to the physicians who analyzed the same transcript.
Strategies in which the agent actively leads the interview and elicits additional information performed significantly better than the passive, user-controlled version. The advantage of the AI was greater particularly in cases where the evaluating physicians expressed lower confidence in their own diagnosis.
An additional validation came from wearable data: conversations that led to a diagnosis of acute infection correlated with detectable physiological changes from Fitbit (increases in resting heart rate, decreases in variability, etc.) in the days prior to the reporting of symptoms.
What does this mean in practice?
The study shows that a well-structured symptom interview led by AI can produce differential diagnosis lists of comparable or superior quality to those created by physicians based on the same verbal material. This does not mean that AI replaces the physician – physical examination, investigations, and complete clinical context remain essential. However, it suggests a real potential for triage and pre-consultation tools that help patients articulate their symptoms better and provide physicians with a richer starting point.
For developers and health systems, the message is clear: the quality of the outcome strongly depends on the agent's ability to ask relevant questions and to complete the medical history, not just to respond to what the user says.
Important limitations
SymptomAI was a research prototype. It is not authorized as a medical device and has not been used to guide the treatment of participants. The reference diagnosis was based on self-reporting, and the population was predominantly composed of Fitbit users. The results should not be interpreted as a demonstration that AI diagnoses better than a physician in a clinical setting.
Conclusion
SymptomAI represents an important step toward the real evaluation of conversational AI in healthcare. By combining a nationwide study with clinical validation and physiological data, Google Research has shown that an agent actively conducting the medical history can achieve a level of quality competitive with physicians based on dialogue. The next logical step will be to test such systems in real clinical workflows, under medical supervision, to see if they can reduce access barriers and improve consultation preparedness.
Main source: Breda et al., "SymptomAI: Toward a Conversational AI Agent for Everyday Symptom Assessment", arXiv:2605.04012 (2026); Google Research blog, 22 July 2026.