As It Has Always Been Done

The most remarkable thing about medical education is how little it’s managed to change. The world races ahead, yet a prospective doctor from ten, twenty, or thirty years ago would feel much at home in today’s classrooms. The dissections, the clinical rotations, the chart-review research papers would all be too familiar, and the pervading sense of perpetual overwhelmA full taxonomy of the emotions available to a third-year medical student would make a fine article of its own. would be closer to home than their own mother’s cooking. But if they paid close attention, they would see the threads of similarity unraveling at the edges. Lymphocyte types are up from about two to somewhere in the low dozens. You can’t treat endometriosis as a psychiatric condition anymore. And, among the failed promises of electronic health records to save the physician from paper drudgery, the attentive student might see something stranger.

An app on a doctor’s phone, attached to a patient’s record, does something miraculous. On its own, it listens to the patient and the doctor interact, across personal history and clinical history, from symptoms to exam findings spoken out loud. It collates silently. It understands the student’s presentation, pulls the pertinent positives and negatives together, and stitches a record of the encounter without so much as a written word from the physician.

The doctor, free from documentation burden, can meet their patient as a person rather than a charting obligation. The medical student sees this technology - magical even to its architects - and is taught to promptly forget it.

“You have to learn to write notes to be a good doctor.” Because what is a doctor other than a system to turn conversations into documentation? And, well, isn’t that how it’s always been?

So the students ignore this wonder. The head goes down, the shoulders slump forward, and the frantic scribbles of the day turn into hours of manual chart editing and frustration at rigid templates.Every hospital system I have seen has pushed for standardized templates. Every specialty still struggles to read the notes of every other specialty, and complains that the templates ruin its own workflow. A remarkable lose-lose-lose.

To me this is a tragic loss of learning. It is a misuse of a student’s time and a paradigm so resistant to change that it robs them of any chance at steering the medicine of tomorrow, relegating them instead to replaceable drop-ins for corporate and hospital medicine.

Is there educational value in note-writing? Absolutely - but is note-writing the education itself or just a convenient reflection of the thought process we seek to inspire? If anyone can make a perfect note from a patient encounter, does that not enable us to seek better ways to nurture knowledge and understanding? Do we want physicians who think, or physicians who document, and if it’s the former, why don’t we pursue that wholeheartedly?

I write this as someone who uses frontier models every day. Fable, 5.6 Sol, Gemma 4, Kimi K3 are entwined in all the work I do. Each day, frontier intelligence manages my calendars, audits my statistics, challenges each conclusion in each research paper I write. In my downtime, they make my dinner reservations. Track my packages. Shuffle ebooks across my devices so I’m always on the same page.

It is becoming unbearably clear that medicine, like most white-collar work, will come to revolve around the use and management of artificial intelligence. The case is nearly self-evident, but let me put it briefly.

Models are extraordinary diagnosticians. In controlled trials, they ask better questions, take better histories, produce better differentials, and make more educated recommendations than the physicians they are measured against. Google’s AMIE, tested against twenty primary care physicians across 159 scenarios with trained patient-actors, was more accurate and won on thirty of thirty-two evaluation axes.Tu et al., “Towards conversational diagnostic artificial intelligence,” Nature, 2025. These are simulated encounters, not live patients, and the gap narrows in messier conditions - I am not claiming the machine is already the better doctor at the bedside.

I am claiming something worse for us. In a randomized trial of diagnostic reasoning, physicians handed a large language model scored no better than physicians with the usual textbooks and search - seventy-six percent against seventy-four, statistically indistinguishable. The model working alone scored sixteen points higher than the physicians did.Goh E, Gallo R, Hom J, et al., “Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial,” JAMA Network Open 2024;7(10):e2440969. The accompanying commentary is worth reading alongside it, and is less flattering to the models than I am being here. The tool outperformed the doctors using the tool. Giving the tool to the doctors changed nothing, and the simplest explanation is that nobody had ever taught them how to use it. That result is this essay, and medical education has responded to it by doing nothing at all.

Models are cheap. A single query costs pennies. A demanding task - one that plans, calls tools, and corrects itself over many steps - runs from a few dollars to well over fifty, which is still a rounding error against the salary it substitutes for. The deeper saving is restraint: a model ordering a test does it in a way that would make a radiologist or a lab manager weep for joy. Clear context for the exam, an exact statement of what it expects or hopes to find, and how any such finding would change downstream judgment. Plenty of doctors show the same discipline. But how many don’t? Decades of neglect of statistical education have led to overtreatment and low-value care - tests, screenings, and procedures that should never have been ordered - costing American patients somewhere between seventy-six and a hundred and one billion dollars a year.Shrank WH, Rogstad TL, Parekh N, “Waste in the US Health Care System: Estimated Costs and Potential for Savings,” JAMA, 2019. Overtreatment and low-value care are one of six categories of waste; the total across all six runs to roughly a quarter of national health spending.

The “human touch” argument fails in a direction you haven’t considered. Yes, all else being equal, people want care from a human being who shows empathy and understands the burden of their condition. All else is not equal. The moment a machine is licensed to provide care at the level of a physician, it becomes a source of medical judgment orders of magnitude cheaper than the equivalent person. How much extra would you pay for a human to diagnose you? For “the human touch”? A hundred dollars? A thousand? What if that human were providing a service measurably worse than the machine’s, and generating more downstream cost, and potential for harm, at the same time?

I ask those questions not to argue that doctors should be replaced, but to show what happens if we insist our value lies in warmth alone. It doesn’t and it never did. What a patient actually needs from a physician is someone who will look at a recommendation - from a colleague, from a textbook, or a hyperintelligent chatbot - and decide whether it is right, and then answer for that decision. Empathy is not a moat. Judgment is the moat, and accountability is the moat, and neither of those survives a doctor who has never been taught how the machine reaches its answer.

A full comparison of the two would take more pages than an infinitely scrolling blog provides. But even if you reject my premise of outright superiority, you must grant that the potential of this tool, in the right hands, places it in the most hallowed company of medical innovation - beside the vaccine, the antibiotic, and the magnetic resonance imager. In any future where artificial intelligence has not made every profession obsolete,A matter of when rather than if, but let us assume the current timelines are dramatically wrong and that we see nothing like superintelligence until the second half of the century. it is all but impossible to argue that AI-augmented physicians are not the immediate future of practice.

Why, then, does medical education treat this as a fad? Eighteen months of preclinical training mentioned artificial intelligence as a throwaway: on the third page of a three-page assignment, as a tool to double-check a differential. An optional step, at that.

For anyone not watching the frontier closely, the gap between that assignment and the actual state of the art is difficult to picture. Far from chatbots, advanced harnesses, long-horizon goal setting, and motivated tool use have turned models from question-and-answer machines into agents that work on tasks for days, research their own answers, check their own assumptions, and spawn other models in swarms beneath them. This year, a frontier model solved nine long-standing problems in pure math, two of which had defeated professional mathematicians since the 1970s.AlphaProof Nexus, DeepMind, 2026. Nine problems out of three hundred and fifty-three attempted - a narrow, search-and-verify success rather than general mathematical insight, as DeepMind itself has been careful to say. They are not yet running companies - the most public test of that idea, Anthropic’s attempt to have a model operate a small vending business, lost money and was talked out of its own pricing by a journalist.Anthropic, “Project Vend,” phases one and two, 2025. Phase one gave away inventory and hallucinated business relationships. Phase two turned a profit in one location and lost more than a thousand dollars in another after a Wall Street Journal reporter persuaded the model to declare an “Ultra-Capitalist Free-for-All.” But note what the failure was. Not stupidity, but judgment, and the absence of a supervisor who understood the system well enough to catch it drifting. That is the job I am describing. And the only reason a single hospitalist cannot yet run an internal medicine ward with such a system is the slowness of medical institutions to move.And HIPAA, which predates generative AI and leaves real gaps - most of all around whether patient data may be used to train or tune a model. It does not, contrary to hospital folklore, forbid running AI on patient data. It requires a business associate agreement with the vendor, which is slower and more contentious than the law’s defenders admit and is nonetheless not a wall. Heads will eventually come out of the sand. Public, financial, and - yes - corporate pressure will force this paradigm on all of us.

So what happens to the students of today? They went through four years of an education designed to make them doctors, and will be delivered into a world of agentic AI. How will an optional chatbot exercise translate to a dozen virtual intelligences, each managing a dozen of its own siblings, each assigned to a patient? Incoming physicians are not merely unprepared for this. They have been abandoned without even the awareness that it is coming, and that it will reshape every part of their working lives.

A curriculum that does not require students to use and manage agentic artificial intelligence is obsolete, and does them a grave disservice. While we flip through flashcards, memorizing which species of mosquito carries which arbovirus, technology and law and finance are racing to train their next generation on steering digital minds with purpose. Bankers and lawyers and software engineers are learning what these systems can do and how to turn the work of thinking machines into outcomes that matter. They are learning to set long-horizon goals, and to catch hallucination, cheating, and misalignment. They are learning to parse the mountain of output these jagged, spiky, alien intelligences produce. Most importantly, they are learning to verify, and to take responsibility for, the inscrutable reasoning that produced the recommendation in front of them.

And future doctors? They’re learning to write notes by hand, as it has always been done. In their downtime, they can study contraindications of drugs no longer prescribed, as it has always been done. Most importantly, they’re being taught to do as they are told and not to question their place in health systems, vast and mighty and terrible and occasionally wonderful - a sharp departure from the academic and forward-looking heritage of the profession.

A medical education not built around cooperation with machines resigns us, the future doctors of America, to be lost in the coming maze. Tossed into an ocean of bits and a deluge of output from opaque systems sold to us by the likes of Epic, we are on a path to surrender our judgment by default. For all the noise about the rising tide of AI slop, what are we doing to teach incoming physicians how not to regurgitate it? Nothing.

Surgeons once refused to wash their hands, and the profession fought antisepsis for a generation before it became unthinkable to practice without it. Evidence-based medicine was a heresy inside living memory. Medicine has never actually been static; it has only ever been slow, and then suddenly not. The tradition of medicine is a tradition of reinvention, and we are currently failing to live up to it.

So let me be concrete about what I want. I want a curriculum where students are handed a frontier model on day one and made to use it under supervision, the way we are made to use a stethoscope. I want assessment that rewards a student for catching a model’s confident, fluent, wrong answer. I want every graduating physician to have argued with a machine about a patient and been able to say why they overruled it.

The students who learn to command these systems will be the clinical leaders of the next forty years, and the ones who don’t will spend their careers taking orders from software they were never taught to question. We are choosing, right now, which of those we are training. We should choose on purpose.