Medical Artificial Intelligence · us
Discussing Symptoms With AI Before the Visit: Google AMIE Completes Physician-Supervised Study in Real Outpatient Care
100 patients completed AI consultations without triggering predefined safety stopping criteria, but zero stops does not mean zero errors. The study brings pre-visit history-taking into real outpatient care while marking the boundary that autonomous diagnosis has yet to cross.
Before entering the examination room, could patients first explain their symptoms to AI, giving physicians more time for assessment and discussion? The AMIE study, a collaboration between Google and Beth Israel Deaconess Medical Center (BIDMC), brought this concept into real primary care clinics. The results, published in The Lancet, suggest preliminary feasibility for pre-visit conversations under continuous physician supervision; whether they can improve care remains unresolved.
This prospective, single-center, single-arm study enrolled 114 patients from April to November 2025. Of these, 100 completed an interaction with AMIE, and 98 also completed the subsequent outpatient visit. Participants were English-speaking adults with a single new, non-emergency presenting complaint who exchanged text messages with the AI up to five days before their visit. Physicians monitored the conversations in real time through video and screen sharing, and transcripts and summaries were provided to the physicians conducting the subsequent visits.
No conversation in the study triggered the predefined safety stopping criteria, but this does not mean the system made no errors. The paper’s abstract reports that supervising physicians identified one “hallucination” and added clinical information during five interactions. The results therefore support the feasibility of supervised conversations; they do not establish that the system could continue to operate safely without physician oversight.
The diagnostic performance figures also require careful interpretation. According to an analysis released by Google Research, using the final diagnosis established through a medical record review eight weeks after the visit as the reference, AMIE included that diagnosis among its top seven possible diagnoses in 90% of cases, among its top three in 75%, and as its first choice in 56%. In other words, listing a broad range of possible causes and selecting the diagnosis on the first attempt are different abilities. Some final diagnoses also reflected physicians’ clinical judgments rather than confirmation through testing in every case.
Pre-visit preparation offered another observation more closely tied to everyday clinical work. Among 44 cases in which physicians had read the conversation transcript before the visit and provided relevant responses, 33 responses indicated that the transcript helped them prepare, and 25 indicated that it might change their approach. This provides a positive signal about the user experience, but it does not yet demonstrate that AI shortened visits or improved patients’ health outcomes.
A gap remained when moving from diagnosis to management. Study ratings showed that physicians’ management plans outperformed AMIE’s in practicality and cost-effectiveness. At the time, the AI could not access electronic health records or perform physical examinations. The absence of prior medical history, physical signs, and information available in person limits its ability to translate recommendations into care for individual patients.
Funded by Alphabet, this study moves medical AI evaluation from simulated settings into real patient care workflows. However, its small size, single-center setting, and lack of a usual-care control group mean it is not yet sufficient to establish clinical benefit or the safety of autonomous use. Google also called for larger trials in its announcement. The key next step is to test whether these pre-visit tools can consistently reduce burdens across different patients and healthcare settings while maintaining the quality of care.