I used to have this machine as a kid - great computer!

True. Also, you don't have to keep consenting patients once they have already consented. I'd worry about automation bias and checking output - better structured summaries and patient access to check the note afterwards would help I think.

I think if you have to consent 3 times (privacy, listening device, algorithm output) and then thoroughly check the output for errors, you start to lose the efficiency gains from AI scribes.

Yes, good to see RCTs in this area now. I think might be feasible to do a trial that uses an ambient AI scribe in a real consultation, then use the summary as a prompt to an LLM to generate a diagnosis. Then compare that to a clinician's diagnosis.

Interesting study but I would argue that this is similar to LLM performance on medical exams rather than real diagnostic reasoning. The cases were presented as written vignettes and LLMs have probably thousands of similar vignettes in the training datasets they can use for pattern recognition.

The course is a well-known online introductory informatics course. The LLMs scored better than as many three-fourths of students taking the course. The big questions for me are their implications for student learning and assessment.

The original paper is at: doi.org/10.1038/s417...

Please add me. Thanks!

Hi Larry! Nice community of digital health folk building on here. :-)