In a 98-patient study at a Boston clinic, Google’s AMIE matched doctors’ final diagnoses 90% of the time, with larger trials still needed.
Google and Beth Israel Deaconess Medical Center reported results from a real-world feasibility study of AMIE, Google’s research diagnostic chatbot, in an ambulatory primary-care clinic. Published in The Lancet on October 8, the study involved 98 patients who consulted the system before urgent-care visits. Supervising physicians monitored the interactions in real time, and Google said none required interruption under the study’s predefined safety criteria.
The researchers reported that AMIE’s differential diagnoses matched the doctors’ final diagnoses 90% of the time. Clinicians said the chatbot’s summaries helped them prepare for visits in 75% of cases and influenced their approach to care in more than half. The findings suggest a possible role for AI in gathering patient histories and organizing information ahead of a clinician encounter, rather than independently directing treatment.
The results are an early feasibility signal, not evidence that a chatbot is ready to replace clinicians or to be deployed broadly. The study took place at a single academic medical center and involved fewer than 100 patients. Google itself said larger clinical trials are needed to assess patient-facing AI at scale. The company’s public summary does not establish whether performance would carry over to other clinics, patient populations or workflows, and the 90% figure refers to whether AMIE’s differential included the doctors’ final diagnosis—not a finding that the AI independently delivered definitive diagnoses.
For healthcare organizations, the operational question is whether pre-visit AI can produce useful summaries without adding new safety, privacy or workload burdens. In this study, clinician oversight was built into the process, and participants’ interactions were monitored. The next meaningful tests will need to examine performance in larger and more varied settings, how clinicians verify AI-generated information, and whether the reported preparation benefits translate into better care or more efficient visits. For now, the paper offers a glimpse of a supervised intake model, not a validated autonomous diagnostic service.