Conversational AI Takes Histories Before Real Primary Care Visits

A conversational AI built by Google has been tested on real patients in a working primary care practice, and it did not need to be stopped once on safety grounds. In a prospective feasibility study published in The Lancet on 8 October 2026, adults booked for urgent appointments at Beth Israel Deaconess Medical Center (BIDMC) in Boston chatted by text with the Articulate Medical Intelligence Explorer (AMIE) before seeing their clinician. A physician watched every conversation live and was ready to halt it. That never happened in 98 completed encounters.

The result is narrower than a headline about "doctor-level AI" would suggest. The study had no control arm, was run at a single centre, and was designed to show that the approach can be done safely and acceptably, not that it improves care. Still, it is one of the first times a diagnostic chatbot of this kind has moved from simulated consultations with actors to real people with real symptoms, and its details show both what such systems can already do and where they fall short.

From simulated patients to a real clinic

AMIE is a research system, not a product. Its earlier evaluations were laboratory studies. The best known, published in Nature in 2025, was a randomized, double-blind crossover study modelled on the objective structured clinical examinations used to assess medical trainees. Patient actors worked through 159 case scenarios by text chat with either AMIE or one of 20 primary care physicians. Specialist physicians rated AMIE higher on 30 of 32 axes, and the patient actors rated it higher on 25 of 26. The authors themselves cautioned that text chat is not how clinicians normally work, and that the results could not simply be carried over to practice.

Simulations leave out much of what makes real consultations hard: patients who are anxious, vague or contradict themselves, who have several problems at once, or whose symptoms settle before they are seen. The BIDMC study was designed to find out whether a patient-facing system could cope with that in a supervised setting.

How the study worked

Enrolment ran from April to November 2025 at the practice's urgent care appointments. Eligible patients were English-speaking adults booked for a visit about a single complaint. Pregnant patients, mental health complaints and anyone triaged as an emergency were excluded. Participants could hold a text conversation with AMIE from home up to five days before their appointment.

According to the March 2026 preprint of the study, AMIE ran on Google's Gemini 2.5 Pro with its reasoning ("thinking") mode enabled. After about 50 encounters the team switched to the faster Gemini 2.5 Flash because responses were too slow. Rather than working through a fixed questionnaire, the system kept an internal running state: a summary of the patient, a working differential diagnosis, a list of information gaps and a draft management plan. It moved through phases of intake, history-taking, checking its diagnostic hypotheses, explaining its assessment and wrapping up. It chose follow-up questions based on what it still needed to know.

Patients were offered possible diagnoses, framed tentatively and with disclaimers, along with next steps to discuss with their clinician. The treating clinician received the transcript and an automatically generated summary that included the differential. AMIE's draft management plan was recorded for research only and was not shared with patients or clinicians.

Supervision was central to the design. A board-certified internal medicine physician monitored each conversation in real time and could stop it under four prespecified criteria: immediate concern for harm to the patient or others, significant distress caused by the interaction, potential clinical harm identified by the supervisor, or a patient asking to end the session.

What the study found

Of 114 patients enrolled, 98 completed both the AMIE conversation and their physician appointment. (The March preprint and the trial registration, NCT06911398, give a figure of 100 participants, and the preprint also analysed 98. The summaries available to BiomedEdge do not explain how the published enrolment figure of 114 reconciles with those counts.)

• Safety stops: none under the prespecified criteria.
• Errors: supervisors noted one hallucination and added clinical information in five conversations. The preprint gives one example of an error that was corrected: AMIE stated that a past operation was scheduled for a date in the future.
• Conversation quality: clinical evaluators rated conversations favourably in 87-100% of cases across 17 criteria. Patients' favourable ratings ranged more widely, from 48% to 96% across 16 criteria.
• Attitudes to AI: patients' attitudes towards AI improved after the conversation and stayed higher after they had seen their clinician. Concerns remained about confidentiality and whether the chatbot could be trusted to be honest.
• Clinician use: clinicians reviewed the AMIE output before 44 visits. They found it helpful for preparation in 33 (75%) and said it may have changed their approach in 25 (57%). One clinician judged an interaction somewhat harmful, because a patient may have been made anxious by seeing lymphoma in the differential.

The preprint also reported diagnostic and management results, measured against a final diagnosis set by blinded internists on chart review eight weeks after the visit. The final diagnosis appeared among AMIE's first seven candidates in 90% of the 98 analysed cases, in its top three in 75%, and as its first suggestion in 56%. In blinded comparisons, the overall quality of AMIE's differentials was similar to that of the patients' clinicians. Its management plans were rated similar for appropriateness and safety. The clinicians' plans were rated significantly better for practicality and cost-effectiveness. These figures come from the preprint; the summaries of the peer-reviewed paper available to BiomedEdge did not report them, so they may differ in the final published version.

Why the practicality gap matters

The management finding is easy to pass over, but it points to a structural limit rather than a flaw that a better model would simply fix. AMIE worked only from what the patient typed. It had no access to the electronic health record, no physical examination, no knowledge of local services, and no sense of what the patient had already tried or could afford. The clinicians had all of that. The authors attribute part of the gap to this difference in context.

That has implications for how such tools might be used. History-taking and drafting a differential before a visit are tasks where a patient-reported account is most of what is available anyway. Deciding on a plan is where the clinician's knowledge of the patient and the health system counts for most. The study's design, in which AMIE's plan was never shown to anyone, reflects that division.

Limitations

The authors and outside commentators list several reasons for caution.

• No comparison group. Without a control arm, the study cannot say whether AMIE changed visit length, diagnostic accuracy, patient outcomes or costs. It was not powered to.
• Heavy supervision. A physician read every conversation as it happened. That is not a model that would scale, and the zero-stop result applies to a supervised system, not an autonomous one.
• A selected population. One academic centre, English speakers only, single complaints, and patients willing and able to use a text chat. People who are more engaged with technology may have been more likely to take part. The preprint reports that 68% of analysed patients were women.
• Uncertain reference diagnoses. Many urgent complaints resolve on their own, so the final diagnosis used to score the differential was sometimes presumptive.
• Being watched. Patients and clinicians knew they were part of a study, which may have affected how they behaved.
• Industry involvement. The work was a collaboration between BIDMC and Google, with Google-affiliated authors, and was funded by Alphabet.

An accompanying Lancet Comment, published the same day, framed the open question as what happens after the AI speaks: how patients and clinicians act on a diagnosis or other clinical output from a chatbot. On that, the evidence is still sparse. The one clinician who judged an interaction harmful, over a cancer appearing in a patient's differential, is a small example of the problem. Showing possible diagnoses to patients before a clinician has seen them may be informative, alarming, or both.

What to watch

• Controlled trials. The next informative step is a randomized comparison of AI pre-visit history-taking against usual care, with outcomes such as diagnostic error, visit time, follow-up and patient harm.
• Less supervision. Whether safety holds when monitoring moves from continuous real-time oversight to sampling or after-the-fact review, which is the only form that would be practical at scale.
• Integration with records. Whether giving a system access to the health record narrows the gap in practicality and cost-effectiveness, and what that means for privacy and consent.
• Broader populations. Multi-site studies including non-English speakers, people with several problems at once, and patients with less digital access.
• Regulation. A patient-facing tool that suggests diagnoses would raise questions about how it is classified and regulated as software, which this research study did not need to address.

For now, the BIDMC study supports a limited claim: under close physician supervision, a conversational AI can take histories from real primary care patients without triggering safety stops, and most clinicians who read its output found it useful. Whether it improves care, and whether it can run safely with less oversight, remains to be tested.

This article is for information only and is not medical advice. Treatment decisions should be made with a qualified clinician.

Sources: The Lancet (Brodeur et al., 2026); arXiv preprint 2603.08448; Medical Xpress / BIDMC; News-Medical; ClinicalTrials.gov NCT06911398; Nature (Tu et al., 2025)