Ouyang: AI Medical Scribes May Blunt Doctors' Reasoning

An emergency-medicine doctor spent months letting AI scribes listen in on her patient visits and write the chart notes for her, and the essay she wrote for The New York Times Magazine is the most useful thing I have read on how ambient AI actually changes the work, rather than how vendors say it will. Helen Ouyang, an emergency physician at Columbia and a contributing writer to the magazine, describes being "eager to try them out" when her hospital rolled the tools out. Then something quieter started happening to her own reasoning.

The line she keeps returning to is that "editing a note I did not create does not demand the same of me as writing a note." At first, she says, the distinction "seemed small." Over time, though, she has "come to see how much of my own thinking had been bound up in the writing process itself." Writing a chart, in other words, was not just paperwork. It was where she rehearsed the differential diagnosis, the pattern-matching, the what-did-the-patient-not-say instinct. When the scribe hands you a clean draft to approve, you skim, and the thinking-by-writing step gets skipped.

Ouyang walks through the history of the chart, from the nineteenth-century rise of hospitals to the 1960s, when Dr. Lawrence Weed at Case Western Reserve University codified the SOAP structure, subjective, objective, assessment, plan, that most clinical notes still follow. The point is not nostalgia. It is that the structured note was designed as a cognitive scaffold for the doctor, not just a record for the chart, and generative AI is quietly stripping that scaffold out.

The reporting also notes that about a third of AI transcriptions contain errors, which is the accuracy caveat clinicians have to sit with even when the tools save real time. The honest limit of the piece is that it is one physician's account. It does not try to answer how many hospitals have gone all-in, what patients think about being recorded, or whether junior doctors trained on AI-drafted notes will develop the same reasoning muscles as the generation before them.

That last question is the one worth watching. If ambient scribes are as sticky as adoption suggests, the practical fight over the next few years will be less about raw transcription accuracy and more about workflow design. Do the tools force the clinician to actually think through the note, or do they optimize for speed and quietly deskill the people using them?