“One of the first doctorly tasks for medical students is learning to write the clinical note. Before they have even finished their anatomy and physiology courses, they are sent to patients’ bedsides with a teaching physician and asked to come back with a written account of what happened there: what the patient said, what was observed and, in due course, as they gain experience, what diagnoses might explain the patient’s symptoms and what might be done to treat them. When I was in med school, this work felt like an earned, almost ceremonial initiation, one of the humble rituals through which a student starts to act and feel like a doctor.
We were there to learn, and without caseloads of our own to manage, we sure had the hours to write these notes. But as we matured into resident physicians, then attending physicians, that time no longer felt available to us, and the production of notes no longer felt like a privilege. They felt tiresome and detached from what medicine was really about: talking with patients, thinking through their symptoms and, hopefully, making them better. Many of us began to see the note as something that had very little to do with — and maybe even got in the way of — treating patients. If we reclaimed all those hours, couldn’t we take better care of people?
So, when the hospital where I work as an E.R. doctor debuted artificial-intelligence assistants — they listen in the background through our phones as we meet with patients, then churn out polished, in-depth notes — I was eager to try them out. I had already heard about their potential, having spoken a couple of years earlier to the chief executive of a leading A.I. medical-scribe company. As he described his product, I imagined the doctor I would become: a better, more attentive healer, someone able to unhurriedly devote more time to patients at their bedsides instead of furiously typing at her computer. (Selfishly, I also envisioned improved interactions with my own physician, whom I see once a year, always with her eyes glued to her screen as her fingers click away on a keyboard.)
This is also the vision that the makers of A.I. scribes are selling to medical facilities. Start-ups have proliferated, raising nearly $5 billion since 2019 and promising to both relieve clinicians from burnout and provide health systems with greater efficiency. For a study published last year, every one of the 43 nonprofit health systems in the United States that completed the survey reported that it was developing, piloting or already deploying A.I. scribes. While of course not every physician will choose to use this tool — though institutional pressures could someday make it feel more like a requirement than an option — the rate of adoption has reached as high as 80 percent in some clinics and departments.
Adam Rodman, a Harvard Medical School professor and the director of A.I. programs at the Carl J. Shapiro Institute for Education and Research, referred to A.I. on a recent podcast as “probably the fastest-adopted medical technology of all time.” For all of the exhilarating possibilities A.I. brings to medicine — diagnosing rare diseases, predicting individuals’ risks, discovering elusive drugs — one of its strongest footholds in health care today is its ability to eavesdrop on conversations and quickly generate notes from them.
This process involves more than merely transcribing a recorded interaction between doctor and patient. An A.I. scribe extracts and synthesizes the information from that conversation — sometimes offering its own take as well — and produces a document in the structure and tone of a clinical note.
The note, I must say, comes out beautiful! No typos, proper grammar, complete sentences. Many doctors are delighted. I have been too. After all, wasn’t this always A.I.’s great seduction? To transport us away from all the boring, tedious busywork? Only later did it occur to me that something might also get lost along the way.
At least for now, though, trialing A.I. to write notes offers maybe the safest way to broadly introduce this technology into health care — as a low-risk application, far removed from the much more frightening prospect of letting machines make life-or-death decisions about treatment. No wonder A.I. scribes are spreading so quickly. Health systems, after years of hearing doctors complain that the computer was coming between them and their patients, are finally investing in something that seems to promise a way back.
The modern chart began to emerge around the turn of the 19th century, as medicine moved into hospitals, where patients were under observation day after day, their illnesses followed over time. Early American hospital records consisted mostly of admission and discharge books, with more administrative information about patients than clinical details. Eventually, hospitals began preserving fuller case histories, retrospectively copied by scriveners from physicians’ private notebooks. By the end of the century, New York Hospital, one of the country’s first hospitals, advised physicians to write at the bedside “with such care and in such a manner” that their notes could become part of the permanent institutional record.
This documentation served the hospital’s record-keeping needs more than those of the individual patient. It wasn’t until 1907, at the Mayo Clinic, in Rochester, Minn., that the chart started to be integrated more directly into the patient’s care. Around the same time, nursing observations, test results and other clinical information also began finding their way into the medical record. Henry Plummer, an endocrinologist at Mayo, and his assistant, Mabel Root, created a dedicated record for each patient, gathering a person’s medical history in one file rather than leaving it dispersed across separate ledgers and admissions documents. The patient, not the ward or the single hospital stay, became the organizing principle. A chart now followed a person across visits and time; if the patient returned, another doctor could pick up the thread of his or her care.
The notes themselves, however, could be chaotic. Lawrence Weed, a physician and research scientist, saw any disorder in doctors’ notes as disorder in thought. He criticized writing that concealed how a doctor had moved from symptom to diagnosis to plan. Weed’s response, while he was a professor in the 1960s at what is now Case Western Reserve University, was to put forward the methodical structure that became the basis for what’s now known as the SOAP note still used today: subjective, for what the patient reports; objective, for what the doctor observes or measures; and, arguably most important, the assessment and plan: what the doctor thinks is happening and what she intends to do about it. This discipline pushed physicians to bring order to their thinking and make their reasoning visible on the page.
For decades, these notes were clipped to patients’ charts at the feet of their beds, or they filled big binders at the nurses’ stations. Then electronic medical records came along, as well as the computers that facilitated them; as this innovation became widespread in the late 2000s, it transformed clinical notes once again. My own development as a doctor coincided with this transition in American medicine. For several years, I bounced back and forth between paper and electronic notes, both jotting and typing, depending on the protocol at the hospital where I happened to be working. As someone who first got a computer in high school, I felt mostly at ease with the electronic medical record and didn’t mind when it fully took over. But several older, much wiser colleagues who were two-finger typists were unable to adapt and had to stop seeing patients. Medicine lost some really good doctors during that time.
Typing offers at least one undeniable advantage, though: It removes the hazardous potential of doctors’ handwriting, which is often as illegible as their signatures. A scribbled “hydroxyzine” could be mistaken for “hydralazine,” whose use might cause a plunge in blood pressure rather than the soothing of an itch; a hurried “1” could look like “7” and lead to a dangerous overdose. Whatever frustrations the electronic medical record later brought — copy-paste clutter, inbox overload, endless clicks — it nonetheless made the note legible and easy to share. And even as the note moved from paper to screen, the words themselves were still composed by doctors.
When I first began using A.I. scribes, I regarded them as the natural next step in this long evolution of the medical record. I had not yet thought to ask whether changes to the note’s structure or the medium in which it was produced were a different sort of advance from letting a machine write the note. Instead, I was focused on all the notes I had to juggle during a busy, exhausting shift and the relief A.I. scribes would provide. As long as the note captured my thinking in some way, even if the words were not my own, what difference did it make how it came to be drafted?
In the E.R., interactions between patient and doctor rarely unfold neatly and logically. A patient starts with a description of chest pain, circles back to a fever, remembers a medication he stopped taking, mentions a family history of heart disease only after his wife prompts him, then adds one more symptom as the encounter is wrapping up. But the A.I. scribe converts that wandering, stuttering exchange into a tidy clinical synthesis — a miracle, it seems! The messy human sprawl of the visit returned to me clean and perfectly punctuated in less than a minute.
For a while, after I started using the scribe, my role in creating the medical note seemed simple: read the draft, correct what was wrong, sign it. So fast, so easy.
But then the process began to discomfit me, as I slowly realized that editing a note I did not create does not demand the same of me as writing a note. I was no longer thinking through the interaction with a patient and the meaning of the information I had gathered from the visit. I was checking a version that had already been made for me.
This distinction seemed small at first. The notes had been generated from my encounters, from the things that were said in the exam room. And I was still revising and proofreading the A.I.’s output afterward. Over time, though, I have come to see how much of my own thinking had been bound up in the writing process itself.
Daniel Kahneman, the behavioral psychologist and Nobel laureate in economics, distinguished between two modes of judgment: the quick, intuitive impressions that arise almost automatically and the slower, more effortful reasoning that questions them. A doctor must work in both. Often, I begin to sense a diagnosis as soon as I walk into a patient’s room, before I can fully explain it; when I write my notes, I tend to slow down to do my more deliberate thinking.
The process of note-writing helps me formulate my medical decision-making and then check whether it really holds up. If I find myself trying too hard to explain away a patient’s symptoms, trying too hard to dismiss a worrisome vital sign, trying too hard to fit an abnormal result into a reassuring narrative, then maybe I need to revise my conclusions. Have I not ruled something out as convincingly as I first thought? Am I tethering myself too quickly to a diagnosis? The patient might need another question, another exam, another test. You have to convince yourself as much as anyone else.
When that cognitive labor is offloaded to a machine, I’ve come to see, my job shifts. Even when I try to speak my reasoning aloud for the A.I. scribe, I am still doing something different from writing the note myself. I am no longer using the note, sentence by sentence, to think through the case in my own words, to decide what to emphasize, what to soften — or, as I’m writing, to identify when my reasoning strains. And unlike when I dictate a note, I can’t watch my own phrasing appear on the screen in real time. With the A.I.-generated note, I am instead auditing afterward. I am playing a version of “Where’s Waldo?” — What’s missing? Has this note gone astray, and if so, where? — and it’s a search made all the more difficult because the A.I.’s draft arrives fluent, confident. It sounds so right.
That cognitive shift does not happen the moment the A.I. scribe delivers a note. It begins in the exam room. Because I know A.I. is recording, I stop listening in the same way. Before A.I. scribes arrived, I would outline a story in my head as a patient talked, fitting the pieces together so I would know what to ask next. In the scribe’s presence, that work is deferred. Let the machine do it! The mind drifts.
In E.R. settings, patients handed off from one doctor to another when shifts change often feel less known to us; an inherited patient can come already filtered through someone else’s account. With A.I. running in the background, something similar happens: Even when I’m the first doctor to see someone, those patients began to feel, strangely, less my own.
That altered feeling in the exam room troubles me for another reason: The machine is eavesdropping. I am not so worried that the audio might surface in a courtroom years later. (We have been told the recordings are stored for three months.) What unsettles me is that the interaction no longer feels quite like what medicine, at its heart, is supposed to be: an intimate exchange between two humans. The hospital room should be able to hold fear, shame and guilt, to shelter the things people say in the depths of illness. I have listened to mothers admit they resent their babies; partners confess infidelity; cancer patients say they are ready to die but loved ones refuse to let go; a subway conductor grieve the man who fell onto the tracks in front of his train. When every word is being recorded, no matter that the recording stays confined to a machine, something feels diminished. There is, or ought to be, something close to sacred in such moments.
The practice of medicine also depends on another kind of exchange: the one between doctors. Last year, a colleague had a follow-up visit with a patient I had seen. Reading my note, she texted me afterward: “I felt your anxiety.” That was what I had hoped that patient’s note would transmit: not just my conclusions but also the pressure behind them — what worried me, where I still felt uneasy. The note was providing what Weed believed the medical record should: “a communication for all time.”
Now, when I read some notes generated by A.I. for other doctors, I am less sure about what my colleagues really thought. Bloated notes can turn a straightforward matter into an opus. A.I. can give as much attention to a stubbed toe as to complex rheumatologic findings and unusual surgical complications. Notes can be well composed without necessarily being clarifying. Often, I find myself skimming, trying to tease out what concerns might lie beneath all the lovely prose. What, exactly, was the doctor trying to say here? What, in fact, was she thinking?
Today, if they want, patients can also read these notes, instantly on an app on their phones — and sometimes they do, even while still lying on the stretcher in front of you. This can occasionally be awkward, though overall it is probably a good thing.
Even so, medical language can confuse or cause alarm. Perhaps A.I. could translate doctors’ notes automatically and make the jargon and shorthand more understandable. I recall a patient who was upset by a note in which another doctor, describing her symptoms, wrote “+anorexia.” The doctor meant poor appetite related to her recent illness; the patient thought she had been diagnosed with an eating disorder, anorexia nervosa. And doctors’ notes have long carried their own codes and biases; A.I. might avoid such stigmatizing expressions.
Currently, though, it’s not foolproof. Recently, a colleague lamented that it had introduced a new offense, calling her patient “an alcoholic.” (At least doctors can conveniently blame the machine now: A.I. did it, not me!)
Just a few years ago, it seemed that the business of medicine was catching on to the fact that rigid, templated notes didn’t necessarily improve the care patients received. The dreaded “review of systems” — a checklist for symptoms we had to go through for every organ in a person’s body — stopped being required for billing purposes. Nor did we have to catalog every feature of, say, someone’s abdominal pain just to be fully reimbursed for having treated it. The focus returned to our medical decision-making, the section of the note doctors long valued the most. Why, then, are we now letting A.I. take over this task?
Its rollout might offer a possible answer. When A.I. scribes were introduced at my hospital, a video circulated that showed several colleagues championing the initiative. Toward the end, one physician listed the selling points: increased productivity, increased revenue and — coming last — improved care. Perhaps the sequence meant nothing. But now I wonder whether the note is being turned over to A.I. to help doctors do better by their patients or to pursue efficiency in yet another area of medicine.
So far, studies have found that doctors who choose to use A.I. scribes experience a reduced burden when it comes to documenting their work and feel less burnout. Many clinicians who rely on the technology also report that it substantially lightens their cognitive load, a temptation I know myself: I too have been drawn to it, particularly when I’m tired, because, well, it makes my job feel less taxing — in the moment, anyway.
But what remains largely unstudied is how A.I. scribes affect cognition in a deeper sense: how outsourcing the note to them erodes the ways doctors think and alters the process by which they decide what matters. The danger is not only that A.I. might write something wrong. It is that, in writing something fluid and plausible, A.I. may make it easier for us to offload some of the reasoning medicine requires, to do less of the painstaking work of sorting through uncertainty ourselves. That loss, I fear, will eventually show up in the clinical care patients receive.
On certain afternoons, in my early years at med school, preselected patients — motivated, I suspect, by the boredom of lying around all day in their hospital beds — surrendered to us students. We timidly repeated meandering questions (patients no doubt wondering why they were being asked the same thing again) and clumsy exams (students wondering why we couldn’t hear through our stethoscopes before realizing the wrong side was touching the patient). Still, we managed to collect precious material in the little notebooks that we stuffed into the pockets of our white coats. We then did our best to piece together our scribbles into a coherent clinical note, struggling on the page to make sense of what these patients had just told us. Though I didn’t know it at the time, it was then that I was starting to think like a doctor.
What proved most thorny were the real, live persons in front of us. Unlike their textbook analogues, these patients didn’t stick to how their diagnoses said they should feel. They had too many symptoms. Or too few. Or they were described in terms that didn’t match what we had memorized. Why did these symptoms point toward one diagnosis and not another? How to explain the signs that seemed off in some way? Our syntheses were often thin or flat-out wrong; we knew too little medicine for it to be otherwise. But that was not really the point. The note was where we began learning how to take a scatter of subjective descriptions and bodily findings and make meaning out of them.
We were taking part in a tradition that William Osler, a founding professor of the Johns Hopkins medical school, helped establish in the late 1800s, in which a medical education was expected to extend beyond the lecture hall. Students were sent to the bedside early in their schooling, where they were told to “observe, record, tabulate, communicate.” Writing was an important part of that apprenticeship, turning what students heard and saw at the bedside into instruction. As the education scholar Janet Emig wrote in an influential 1977 essay, “Writing represents a unique mode of learning — not merely valuable, not merely special, but unique.”
A student’s note is rarely elegant. But it shows the conversion that medicine requires: from describing to understanding to judgment. It demonstrates what has been noticed and missed, what has been overvalued and underestimated. “The student’s way of organizing the data should reveal to the teacher whether or not the student writing the history sees the whole story,” Weed wrote. When I work with students in the E.R., their notes, imperfect as they are, tell me how a future doctor is learning to think.
The next generation of doctors is coming of age in a world where A.I. can summarize research, answer clinical questions, make diagnoses. I wouldn’t suggest that students and trainees should be insulated from A.I. entirely. The tools may be shiny and new, but the habit is not: Doctors have always used reference works, from pocket manuals to door-stopper textbooks.
In the E.R., I won’t hesitate to consult A.I. platforms, such as OpenEvidence, a professional chatbot that is trained on peer-reviewed medical literature; more than ever before, they provide me with information that is current, easier to find and personalized for patients. In these ways, A.I. can help me return to the bedside with sharper judgment, armed with the latest evidence.
But that is different from outsourcing the written account in which a doctor — or a student who will become one — thinks through a patient. As far as I know, medical students are not using A.I. to write their notes. I worry, though, that as the profession grows more comfortable treating the note as something A.I. can handle, the exercise of writing one may begin to seem expendable.
Which is why an email I received several months ago felt so consequential — and so disheartening. Third-year medical students rotating through our E.R., it read, would no longer be required to write notes during their shifts.
Maybe I will eventually figure out how to use A.I. scribes without letting them shortcut my clinical reasoning. Maybe the editing process, if done carefully enough, can spur some of the thinking that writing once demanded. Or maybe I will simply be swept along in their uptake, as A.I. scribes become less a choice than an institutional norm.
In medicine, what becomes routine inevitably becomes part of what trainees understand doctoring to be. If note-writing recedes, any loss might not be obvious at first, obscured in part by the immediate rewards of efficiency. Later, however, we might realize how much the making of a doctor depended on that work.
Helen Ouyang is a physician and associate professor at Columbia University and contributing writer for the magazine. She is also a fellow at the Type Media Center.” [1]
1. How A.I. Might Change the Way Doctors Think. Ouyang, Helen. New York Times (Online) New York Times Company. Jul 1, 2026.
