AMIE, Google’s experimental system, spoke with real patients before their medical appointment in Boston. The study published on October 8, 2026 shows that such an exchange can be integrated into a supervised consultation. It does not demonstrate better patient health or the possibility of replacing the physician.
The change is due to the evaluation field. Google and Beth Israel Deaconess Medical Center researchers left the only simulated consultations. Their prospective study was conducted between April and November 2025 in a primary care structure. It included 114 people. Among them, 98 completed both the exchange with artificial intelligence and their consultation.
These results, which were released in October, do not constitute a new discovery in November. They allow us to examine a specific question: what is an automated medical conversation worth when it precedes real care?
AMIE prepares an appointment, without replacing it
Participants had already booked a consultation for a one-time problem. The triage of the clinic had ruled out situations requiring emergency care. They then exchanged in writing with AMIE, from home, until five days before their appointment. The device collected symptoms and history, and then presented assumptions for discussion with the caregiver.
A summary and transcript could be sent to the consulting professional. The experiment was not to let a catbot decide on a treatment alone. It added a preparatory stage to a continued medical journey.
The protocol selected English-speaking adults, who had already been followed in the facility and were able to use its digital device. Known pregnancies and psychiatric reasons were excluded. These criteria limit the scope of the results: they do not allow to conclude for all patients or clinical situations.
No safety stop, but no zero error
A doctor monitored every conversation in real time. He could interrupt the exchange in the event of risk, significant distress or the participant’s request. No judgment based on these criteria was necessary. However, supervisors noted hallucination, i.e. unsubstantiated information, and provided clinical clarification in five exchanges.
The absence of interruption therefore does not mean that there is no error or useless oversight. It describes the operation of a framed device, with a professional available to intervene. Removing this professional would change the conditions under which the result was obtained.
The institution also reports a case considered somewhat harmful by a caregiver. He feared that the mention of a lymphoma among the hypotheses might have caused anxiety. The episode recalls that a medically possible list can have effects on the person receiving it.
Percentages require their denominators
Professionals responded to the questionnaire after 60 of the 98 consultations. They had consulted the AMIE report prior to the appointment in 44 cases. Of these 44 returns, 33 reported assistance in preparation, or 75%. In 25 cases, they felt that the exchange had been able to change their approach.
These responses measure perceived usefulness. They do not prove that three quarters of all patients received better care. They also do not quantify a systematic time-saving or cost reduction.
Another result requires the same precision. In the exploratory analysis presented by the researchers, the diagnosis finally chosen was among the first seven assumptions of AMIE in 90% of cases. It ranked first in 56% of cases, and among the top three in 75%. Presenting the first digit as a single correct diagnosis rate would be misleading.
A comparison still far from an efficiency test
The evaluators appreciated the diagnostic proposals and management plans in an anonymized manner. Physicians achieved better results on the practicality and cost-effectiveness of plans. AMIE had neither a complete medical record nor a physical examination. These differences in context are important in interpretation.
The study remains monocentric and has no control group. It does not compare two random routes, with and without the tool. It therefore does not dissipate its effect on complications, diagnostic errors or workload. High satisfaction and fluid conversation alone are not clinical improvements.
Alphabet funded the work, in which its researchers participate. This link is stated and must accompany the reading of the results. It makes their reproduction particularly important by other teams, in different populations and organizations.
The next steps remain broader comparative trials. They will need to measure patient outcomes, costs, confidentiality and the level of monitoring required. Transposition to Francophone or Arabic-speaking consultations, particularly in Lebanon, is not established by this study conducted in English.
Photograph: Ajay Suresh, BIDMC, November 6, 2025, CC BY 4.0. Archive, proportional reduction and Libnanews dressing.





Comments