On July 22, 2026, Google Research published SymptomAI, a conversational agent that conducts an end-to-end patient interview and then produces a differential diagnosis. Unlike a symptom checker that asks a fixed questionnaire, SymptomAI holds an open-ended conversation, asking follow-up questions as a clinician would before proposing a ranked list of possible conditions.
The evaluation was unusually large for this kind of system. Google ran a randomized national-scale study with 13,917 consenting participants, each of whom interacted with one of five SymptomAI agent variants built on Gemini 2.0 Flash and differing in prompting strategy. Participants received a differential diagnosis from the agent, then reported back two weeks later with the diagnosis their actual healthcare provider had given. Separately, board-certified clinicians reviewed the conversation transcripts blind and ranked both the SymptomAI differential and a clinician-generated differential.
The blinded experts rated SymptomAI’s differential as the best available in 53.3 percent of cases, more than half the time. SymptomAI also showed higher top-5 accuracy than the clinician baselines at identifying the self-reported real-world diagnosis. Notably, its largest advantage appeared on the low-confidence cases, the ones where the reviewing clinicians themselves felt least certain. Google also reports that wearable biosignal data from participants showed physiological trends consistent with the respiratory infection diagnoses.
For business and technical leaders, the point is not that an agent should replace a doctor - Google frames this as symptom assessment, not treatment. The point is the shape of the result: a general-purpose model, on a cheap tier, prompted well, beat a professional baseline on a structured judgment task, and beat it hardest where the humans were least sure. That is the pattern to look for when deciding which internal expert workflows are worth piloting.