← Back to Home

Discussing Symptoms with AI Before a Visit: Google’s AMIE Study Published in The Lancet Demonstrates Feasibility Under Supervision

98 patients completed an AI conversation and a subsequent clinic visit, and some physicians found the records helpful in preparing for visits. But with supervision throughout and no control group, the study cannot yet answer whether AI can improve clinical outcomes.

By SURL BioNews

Before patients enter the examination room, could artificial intelligence first ask about their symptoms to help physicians understand their concerns more quickly? Google’s conversational medical AI system, AMIE, has undergone this test in a real primary care clinic. Google Research confirmed on October 8 that the study had been published in The Lancet: the findings support the preliminary feasibility of conducting pre-visit conversations under real-time physician supervision, but whether this can improve the quality of care remains unanswered.

This prospective, single-center, single-arm study, conducted in collaboration with Beth Israel Deaconess Medical Center in the United States, recruited English-speaking adults who had scheduled clinic visits for new, non-urgent symptoms. Patients conversed with AMIE by text up to five days before their visits; supervising physicians monitored the process through live video and screen sharing, intervening when necessary. Conversation transcripts and summaries were then provided to the treating physicians as materials to prepare for taking the patients’ histories.

The abstract of the formally published study states that 114 participants were enrolled between April and November 2025, of whom 98 completed both the AI conversation and the subsequent clinic visit. None of the conversations triggered the predefined safety stopping criteria, but supervising physicians still recorded one AI hallucination and supplied additional clinical information in five interactions. Therefore, “no stopping” cannot be interpreted as “no errors”; this safety record must also be understood in the context of continuous human oversight.

Whether the AI-organized information was useful depended on how physicians actually used it. Post-visit physician questionnaires were obtained for 60 of the 98 cases. In 44 of those cases, physicians had read the transcripts before the visits; in 33 of these 44 cases, they considered them helpful for preparation. This provides a positive signal for the workflow, but it does not yet support an inference that AI shortened visits or reduced missed diagnoses.

A preprint of the same study released in March reported that 100 adults completed the AI text conversation; Google’s research explanation further noted that 98 of them attended the subsequent clinic visit, with the two figures reflecting different stages of completion. The preprint also reported that, using diagnoses from a medical record review eight weeks after the visit as the reference, AMIE’s list of possible diagnoses included the final diagnosis in 90% of cases, with it appearing among the top three in 75%. These figures describe the performance of candidate diagnosis lists and cannot directly be treated as the success rate of AI independently establishing a diagnosis.

Blinded assessments in the preprint found no statistically significant differences between AI and primary care physicians in overall differential diagnosis quality or in the appropriateness and safety of management plans; however, the absence of a statistically significant difference does not establish equivalence. Physicians’ plans scored higher for practicality and cost-effectiveness. AMIE could not access patients’ electronic health records or perform physical examinations, indicating that a gap remains between text-based reasoning and comprehensive clinical judgment.

The study was funded by Alphabet and did not include usual clinic workflows as a control. Its clearest achievement so far is therefore demonstrating that a supervised process for gathering information before a visit can function. Larger studies with control groups are still needed to determine whether it improves patients’ health, saves healthcare resources, or applies to other languages and healthcare settings; the current findings are also insufficient to support unsupervised use in diagnosis and treatment.

References

  1. Google Research / The Lancet
  2. Google Research
  3. arXiv