A Google and Beth Israel Deaconess Medical Center study of the AMIE conversational medical AI system was published in The Lancet on October 8, 2026. The small prospective study tested whether a research chatbot could collect patient histories before urgent primary-care visits under live physician supervision. It reported no predefined safety stops in the observed conversations and promising diagnostic suggestions, but its single-arm design cannot establish that using AMIE improves clinical outcomes over usual care.
From March research report to October Lancet publication
Google discussed the prospective research in March, when it circulated an earlier study manuscript, and announced the peer-reviewed Lancet publication on October 8. The journal article describes a supervised feasibility investigation, not a general release of an AI doctor. AMIE stands for Articulate Medical Intelligence Explorer, a research system designed for diagnostic dialogue. The newly published work matters because some earlier evaluations of medical chatbots used trained actors or simulated clinical scenarios; this study involved real patients in a clinic workflow. That step adds useful evidence but does not eliminate the need for large controlled evaluations.
How the real-world study worked
According to Google’s detailed research account, 100 adults completed a text conversation with AMIE before scheduled visits for new, non-emergency complaints. Ninety-eight subsequently attended their ambulatory primary-care appointments at BIDMC. Patients agreed to participate, and physicians monitored AMIE chats in real time with permission to stop a conversation based on predefined safety conditions. The system provided transcripts and summaries to the treating clinicians before appointments. Researchers assessed clinician and patient experiences, the safety of observed conversations and the quality of diagnostic possibilities and management plans. The study was conducted at one academic clinical site and did not randomly allocate comparable patients to an ordinary-care control group.
Results: safety, diagnostic suggestions and clinician use
The researchers reported zero interventions to stop AMIE conversations under their defined supervision criteria. Google says physicians considered the AI summaries helpful for preparation in 75% of cases and that they influenced their approach in more than half. The study’s diagnostic accuracy requires careful wording: AMIE included the final diagnosis among its top seven differential diagnoses in 90% of cases. Its top three contained the final diagnosis in about 75%, while the first-ranked suggestion matched in 56%. Those measures do not mean AMIE independently diagnosed 90% of patients correctly on its first choice. Patients generally rated the conversations favorably, and attitudes about AI improved after participation in the limited study.
Why the performance numbers need context
Separate blinded clinical reviewers rated the overall quality of AMIE’s differential diagnoses and management plans as broadly similar to those of primary-care providers for the evaluated cases. However, physicians performed better on the practicality and cost effectiveness of management plans. AMIE lacked some context available to treating doctors, including access to the patients’ complete electronic health records, direct physical examinations and real-time nonverbal cues. The supervising physicians were an active part of the safety model, so these observations should not be applied to unmonitored chatbot consultations. Researchers also noted that the participants skewed younger than the broader population seen at the clinic, making generalization to all primary-care patients uncertain.
What the findings do not establish
Because the investigation was designed to test feasibility rather than clinical efficacy, it does not determine whether AMIE reduces missed diagnoses, shortens appointments, improves patient health or saves money compared with standard care. Nor does the lack of safety stops prove that unsupervised deployment would be safe in other settings. Larger, controlled studies are needed to quantify benefits, harms, workload effects and performance across different populations and kinds of illness. Google has described other planned or ongoing clinical research, but those projects should not be conflated with the evidence from this 100-person pilot. For now the publication supports cautious investigation of clinician-supervised pre-visit AI history-taking; it is not a recommendation to replace professional medical evaluation.
