Google's AMIE chatbot interviewed 98 patients before urgent-care visits in a Lancet study

The design is simple. The patient types their symptoms to AMIE, the chatbot asks follow-up questions, and the doctor gets the full conversation and a summary before the appointment ^[10]. The headline 90 percent needs the same plain reading. In theneuron.ai's description, it means the doctor's final diagnosis appeared somewhere on AMIE's list of possible diagnoses ^[3]. This is a recall-style score. It credits the right answer for turning up anywhere in the candidate set, however many wrong candidates sat beside it. A longer list catches more final diagnoses. Neither account states how long AMIE's list ran, or where on it the right answer sat. For the 90 percent to carry to another clinic, that list would need to be about as long and its candidates about as specific.

The score was also measured with help. Because that summary reached the doctor first, the final diagnosis was often formed with AMIE's candidate list already in view. We would not expect the same match rate where the clinician works the case cold.

On safety, no conversation hit the study's predefined stopping criteria ^[5]. A human stepped in on six of the 98 conversations ^[13].

theneuron.ai, which reported the result, ran a note the same day on how an AI summary can quietly turn a tested tool into one that sounds available now ^[12].