SymptomAI от Google: диалоговый ИИ-агент оценивает симптомы на базе Gemini Flash 2.0

Google Research and Google DeepMind unveiled SymptomAI on July 22, 2026 — a conversational AI agent built on the Gemini Flash 2.0 model that interviews patients in text about their symptoms and produces a list of probable diagnoses. In a national study with 13917 participants, clinicians rated its differential diagnosis as the best option in 53.3% of cases.

How SymptomAI works

SymptomAI carries on a natural-language dialogue with a person: it asks clarifying questions about complaints and, based on the answers, builds a differential diagnosis with recommendations. In the study, each participant was randomly assigned to one of five agent variants that differed in interview flexibility — from a rigid script to a free-form conversation.

• Published — July 22, 2026, by the Google Research and Google DeepMind teams
• Base model — Gemini Flash 2.0
• Participants — 13917 people, a national randomized study
• Five interview strategies: Base, Fixed Canonical, Flexible Canonical, Dynamic Live, Dynamic Final
• Two weeks later, participants reported their doctor's actual diagnosis via survey

How accurate was the agent?

Clinicians rated SymptomAI's differential diagnosis as the best option in 53.3% of cases in a blind evaluation of transcripts. By the top-5 accuracy metric, experts judged the agent's list to be more accurate than those produced by other doctors, and all strategies with active questioning significantly outperformed the baseline Base variant. The authors' key finding is that much of the diagnosis rests on the conversation itself.

"A significant share of clinical diagnoses can be obtained from a language interview alone," the

Google Research publication states.

What do wearables have to do with it?

For some participants, researchers collected up to 30 days of data from Fitbit trackers recorded before the conversation with the agent. This data shows distinct physiological shifts — in cardiovascular activity, breathing, skin temperature, and sleep quality — that preceded the onset of symptoms. This hints that models may one day be able to link a patient's complaints to objective biosignals.

Where the method's limits lie

The authors explicitly list the limitations. A differential diagnosis is inherently ambiguous by nature, and clinicians in the study evaluated static transcripts, while the agent itself conducted an active interview — the conditions are not equal. SymptomAI cannot see body language, visible signs, or the medical record, and its diagnoses were obtained solely for research purposes and do not constitute a validated clinical assessment.

What this means

Large conversational models have come close to structured history-taking: in a sample of nearly 14,000 people, the agent built on Gemini Flash 2.0 outperformed live doctors on differential-diagnosis quality in half of the cases. This is not yet a replacement for a doctor's visit, but it is a claim that an initial symptom review can be scaled through chat.

Frequently asked questions

What is SymptomAI?

It is an experimental conversational AI agent from Google Research and Google DeepMind built on Gemini Flash 2.0: it asks patients about their symptoms and generates a list of probable diagnoses with recommendations. For now it is a research project, not a medical service.

Can SymptomAI be used to make a diagnosis?

No. The authors emphasize that the agent's diagnoses were obtained solely for research purposes and do not constitute a validated clinical assessment. It cannot replace a visit to a doctor.

What model does SymptomAI run on?

The agent is built on Gemini Flash 2.0. The study tested five interview variants — from a rigid script (Fixed Canonical) to a free-form dynamic conversation (Dynamic Live, Dynamic Final).