The Scribe Writes Fluent Notes. Who Writes the Doctor?

The most consequential change in American medical training this year is not going to happen inside a lecture hall.

Applications opened on the first of this month for the Yale School of Medicine's new Master of Health Science in Medical AI, a hybrid graduate degree directed by Xenophon Papademetris and Allen Hsiao. The inaugural class matriculates in August 2027, and applications are accepted through the following February. The University of Florida College of Medicine put its own thirty-credit-hour Artificial Intelligence in Biomedical and Health Sciences master's programme on the fall 2026 calendar. That is the credentialing story, and it is the version of this month's news that will land in most trade publications this week.

The other change, the one nobody applied for, has been happening across roughly every large US teaching hospital for the last eighteen months. The ambient AI scribe entered the exam room, and the second-year resident quietly began learning a different job than the one her attending learned.

1. What the scribe changes

Read Cleveland Clinic's July 2026 npj Health Systems paper on the enterprise deployment. In four months, they rolled ambient AI documentation to more than 4,000 ambulatory clinicians. On the survey side of the deployment, about 60 percent of Cleveland Clinic clinicians using the scribe said the tool increased their likelihood of remaining in practice, and several told the health system directly that they planned to defer retirement because of it. Kaiser Permanente Medical Group's own numbers a year earlier ran in the same direction, with an aggregate reduction in physician documentation workload equivalent to more than 15,700 hours saved in the first year of use across the medical group.

Those figures are real. Read them as the operational win they are. What they do not measure, because it is not their outcome variable, is what happens to the third-year internal-medicine resident who is now handing the assessment-and-plan to the scribe, correcting a paragraph, and clicking sign.

2. What the trainee-side study is beginning to show

The Yale nonrandomised trial published in JMIR Medical Education in 2026 is the first study I have seen that measures the trainee side of ambient scribe use directly, on medical student objective structured clinical examination notes. Its finding is careful. Ambient scribes can lift the note quality of lower-performing students, but the authors flag overreliance risks that scale with time-in-programme and call explicitly for educational safeguards. Read alongside a pilot study of 48 internal-medicine residents at a US programme who generated close to a thousand notes with an AI scribe over six months, and mapped their findings back to seven ACGME core competencies, the picture holds together. Notes improve. Time is saved. Reflection — the specific cognitive act of prioritising a patient's problem list before writing it down — gets thinner. The Johns Hopkins team, publishing the guardrails paper, wrote it as a call for specialty-society guidance, because current ACGME language does not yet cover the scribe-shaped hole in the training day.

3. The AAMC's numbers, read carefully

The AAMC and AACOM Curriculum SCOPE Survey reported that the share of MD- and DO-granting schools in the United States and Canada incorporating AI into their curricula rose from 53 percent in 2023 to 77 percent in 2024. Seventy-eight percent of schools now offer some form of generative-AI user training in workshops or resources. Seventy-four percent cover ethical use. Fifty-six percent teach prompt engineering.

Those are the highlight figures the AAMC has been publishing alongside its AI Competencies across the Learning Continuum framework, the same framework the Icahn School of Medicine at Mount Sinai invoked when it became the first US medical school to deploy OpenAI's ChatGPT Edu to all medical and graduate students in May 2025.

Now read the same numbers the other way. Twenty-three percent of MD and DO schools were not incorporating AI at all as of 2024. Forty-four percent do not teach prompt engineering. A systematic reading by OnlineMedEd put the sharper point: multiple parallel surveys have found that a large majority of medical students had never sat through a single dedicated hour of AI instruction even as they were already using generative-AI tools daily to draft differentials, summarise papers, and study for boards. The curriculum number and the practice number are moving at different speeds.

4. The counter-argument, engaged

The strongest published counter to my reading is Ke and colleagues in Nature Medicine, "AI-induced never-skilling in medical education", which draws a distinction the field has needed. Deskilling is the loss of a competency an experienced clinician once had. Never-skilling is the failure of a trainee to develop the competency in the first place, because AI performed the task from her first shift onward. The authors' point, on the merits, cuts against a comfortable reading of the ambient-scribe deployment data. The retention gains for senior clinicians and the note-quality gains for junior ones can coexist with a slow, invisible failure of the reasoning-development pathway, because the pathway is not what any of the reported metrics currently measures.

Where I would push back on the paper is on one detail only. The authors are careful to state that direct evidence from clinical training is still absent. That is true. It is also the pattern of every early-stage educational-technology introduction in medicine on the record, going back to the transition to PACS in radiology in the early 2000s and the subsequent literature on radiologist search behaviour. The harms are visible late, and the metrics that would have caught them earlier are always designed after the harm has been demonstrated somewhere else.

5. Two layers that do not talk to each other

The Yale MHS and the UF AIBHS master's are the credentialing layer. They will produce, on the current admissions math, in the low thousands of graduates per year across a decade. The AAMC framework is the curricular layer, and it is doing serious work at the level of institutional policy. Neither of these two layers covers the operational layer where a second-year resident and an ambient scribe are already meeting three days a week on the wards.

At the operational layer, the question the scribe answered was documentation burden. It is the wrong question for a training programme. The question the programme needs answered is this: how does a second-year internal-medicine resident learn to hold a differential in her head after a bedside encounter, and articulate it in a note whose structure reveals her reasoning, when the note is being drafted by a system that has never held anything?

6. What a reconciliation ritual looks like

The Journal of General Internal Medicine guardrails paper proposed a structure that translates cleanly across specialties. The resident writes an independent one-paragraph assessment first, reads the scribe's draft only after, and signs a reconciled note that shows the divergences and the reasoning behind each edit. The workflow adds time back to the resident's day. It removes some of the documentation-burden savings the scribe was purchased for. It preserves the reasoning-development pathway the AAMC framework is aiming at, and it makes the never-skilling risk measurable in a way retention surveys never will.

That is not universal practice. It is not close to it. What is universal is that the scribe is in the room. The reconciliation ritual is the design decision still open on almost every teaching service in the country.

7. Whose Monday morning

For the attending running a morning walk-round in a US teaching hospital this week, the tension is immediate and small. She has a scribe running in the exam room. She has an intern beside her who will graduate to a residency in nine months, where a different scribe will be running. She has an AAMC framework on her desk. She may, if she is at Yale or UF or Mount Sinai, know a colleague thinking about a master's application this month.

None of those objects tells her whether to ask her intern, before they read the scribe's draft together, to write two sentences of her own about what she thinks is going on with the patient. That decision is hers. It is also the whole training programme, held in a two-sentence habit.

8. An open question

Ambient scribes are being purchased at the health-system level, with retention and burnout as the outcome variables. AI degrees are being launched at the school-of-medicine level, with credentialing and workforce leadership as the outcome variables. The AAMC framework sits between them, describing what a graduate should know when she leaves. None of those three objects is the object that will decide whether a resident finishing her training in 2029 can still hold a differential without a machine to draft one back at her.

So who owns that question, and at what meeting is it answered before the operational curve makes the answer academic?

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.