The New England Journal of Medicine has published a medical mystery in which nothing is mysterious, a diagnosis in which nothing is diagnosed, and a triumph of artificial intelligence in which the machine recited the opening page of the chapter. The Perspective is titled "What the Robot Still Can't Do," and it appeared on November 14 under the name of Shantanu Nundy, a primary care physician at a community clinic in Virginia who is also the Chief Medical Officer of Humana.1 The journal's own summary describes the author as stumped.
The patient was a thirty-seven-year-old mother of two who had lived with abdominal discomfort for nearly a decade, punctuated by bouts of pain, nausea, and vomiting. Her examination was normal. Her test for the bacterium that causes ulcers was negative twice. A specialist had passed a camera into her stomach and found nothing. A month of acid suppression had done nothing. At her second visit she told her physician, through an interpreter, that the attacks came on when her infant son fell, was hurt, or cried out. He watched the memory of it cross her face. He reports that he still did not know what to do. So, more or less on a whim, he typed her case into a chatbot he had recently downloaded, and the chatbot answered at once that functional dyspepsia led the list. It recommended a low dose of an old antidepressant. Three weeks later she was much improved, and she told him he was her brother, sent by God.
It is a touching story, and the internet has received it as a technological one. The essay has been passed around with admiration by technology founders, from whom one expects no better, and by physicians, from whom one does.2 The posts have the breathless quality of men announcing that a calculator has solved arithmetic.
Consider what was solved. Dyspepsia is Greek for bad digestion. Functional is the profession's word for a complaint that has no identifiable basis and for which there is no evidence of disease. Assemble the two and the machine's answer, rendered in plain English, is that her stomach hurts and nobody can find a reason. She knew that. She had known it for ten years, and she was the one who told us. Molière wrote this scene in 1673, when his candidate physician explains to an admiring faculty that opium causes sleep because it possesses a dormitive virtue. The faculty applauded then, too.
The label has the further distinction of being the likeliest answer before a single test is ordered. Twenty years ago this same journal told its readers that the evaluation of chronic upper abdominal pain usually uncovers no structural cause, and that most such patients have functional dyspepsia.3 Dr. Nundy's own reference, consulted after the chatbot had spoken, told him the condition affects ten to twelve percent of Americans.4 Here is the sentence that should have stopped the editors. Of functional dyspepsia, he writes that he had "never, in 15 years of practice, formally made the diagnosis," nor knowingly met a patient who had it.5 A primary care physician simply cannot practice for fifteen years without seeing functional abdominal pain. He has seen this condition many hundreds of times. What he had never done is recognize it. Medical students are taught that hoofbeats should suggest horses before zebras. The machine heard hoofbeats, consulted the accumulated written knowledge of mankind, and proposed a horse, and a man who has lived his whole career on a horse farm declared it a revelation.
The patient, moreover, had handed him the answer herself. Her stomach seized when her baby cried. The connection between distress and the gut is among the oldest observations in medicine, familiar to anyone who has ever felt ill before an examination, and it requires neither an engineering degree nor a neural network. Dr. Nundy recorded it as a subtle finding and went looking for help.
The defenders of the term deserve a hearing, and so does the outcome. Functional dyspepsia has formal criteria drawn up by an international committee. A review in the New England Journal of Medicine this January proposes that low-grade immune activation in the upper gut may drive symptoms in some patients, and reports that low-dose tricyclic antidepressants outperformed placebo in a randomized trial.6 Above all, the woman got better, and one improved patient outweighs a good deal of commentary, mine included. I grant all of it, and every concession makes the episode worse. The cure was a generic drug older than her physician, one he says he has prescribed hundreds of times for other complaints. For ten years nothing stood between this woman and a cheap bottle of amitriptyline except a doctor willing to name the most ordinary explanation in the book.
In daily practice, the word functional does the work that psychosomatic did before it became impolite, and the history of the vocabulary is a history of flight. Neurology began with hysteria, retreated to conversion disorder, tried psychogenic, and has lately settled on functional neurological disorder.7 The attacks once called pseudoseizures, and then psychogenic seizures, are to be functional or dissociative seizures under the latest proposal from the international epilepsy league.8 Gastroenterology is a step ahead. Its governing committee concluded in 2016 that functional had itself become stigmatizing and began replacing it with disorders of gut-brain interaction, which is the phrase Dr. Nundy's reference supplied to him.9 Every specialty keeps such a drawer, and the cycle in each is the same. A term is coined to avoid saying what the doctor suspects, patients eventually work out what it means, and a fresh term is commissioned. The epilepsy league's own newsletter reports a study in which patients found the gentler names less offensive only until the names were explained to them.10 The reformers do have a point, since no clear psychological trigger can be identified in many such patients, and the mind is not always the whole story. In this patient, however, it was at least the opening chapter, because her pain arrived when her baby cried. Our grandmothers would have called it a nervous stomach, which in her case was more candid than any of its successors and no less kind.
I mean no slight to the patient. Her pain is real, she is not inventing it, and she is entitled to the truth about it. The truth is a short speech that any competent clinician has given many times: we have looked thoroughly, nothing dangerous is causing this, further testing will not find what earlier testing missed, the nervous system and the gut are in a quarrel that we can sometimes quiet, and here is a medicine worth trying. That speech takes ten minutes and a little nerve. It treats the patient as an adult who can bear accurate information. Each new name is a way of postponing it.
The speech also carries a professional hazard that the chatbot's admirers have overlooked. A physician who tells a woman that her tests are normal, that her pain is bound up with her worry over her child, and that he would like her to take an antidepressant can expect to be told, at length and in public, that he is a gaslighter who dismisses women's pain. The chatbot delivered that identical message, and the same public has pronounced it brilliant. One of those verdicts is wrong. I believe the first is, since the message was true and the treatment helped. A culture that hears candor from a doctor as contempt and candor from software as genius will, in time, get its candor exclusively from software.
The machine itself deserves no abuse. It performed as a serviceable index to the textbook, and I have no quarrel with a physician who looks things up. I look things up daily. The abuse belongs to the scale on which the result has been weighed, and the heaviest thumbs on that scale belong to the men who announce, at regular intervals and with considerable funding, that they are rebuilding American medicine. They have watched these systems do astonishing things elsewhere and assumed they were watching it again. In May, an OpenAI model disproved a conjecture about points in the plane that Paul Erdős posed in 1946, and nine mathematicians, Noga Alon and Timothy Gowers among them, checked the proof and confirmed it.1112 That was discovery. Nobody on earth knew the answer, and now we do. But in this case, every textbook knew the answer, the reference website knew the answer, and the author confirmed it there within minutes. AI did not solve a medical mystery. It simply retrieved a paragraph from a rudimentary textbook, at the level of a first-year medical student. A man who cannot tell the two apart has disclosed how little he knows of the field he proposes to rebuild, and he has done so with a confidence that only that ignorance could sustain. Friedrich Hayek observed that the knowledge on which an economy runs is dispersed among the people doing the work and is unavailable to the planner who would direct them from above.13 The founders have every right to enter medicine, and I wish them paying customers. They should understand that a planner in a hooded sweatshirt is still a planner, and that the most ordinary Tuesday in a primary care clinic currently looks to them like a miracle.
They are fortified by a literature of studies announcing that the machines now outperform the doctors, and the studies repay a close look. The best known handed physicians and a chatbot written vignettes, complete with history, examination, and laboratory results, and found that the chatbot alone outscored the physicians.14 Microsoft converted 304 published case records from the New England Journal of Medicine into a quiz and reported 85 percent for its system against 20 percent for practicing physicians. It obtained the second figure by forbidding the physicians their references and their colleagues, a condition under which Dr. Nundy himself would never have reached functional dyspepsia, and the company concedes that its system awaits testing on common, everyday presentations, which is to say on medicine.15 In each case someone has already done the difficult part, which is to extract a coherent history from a frightened person who does not speak in vignettes. When researchers at Oxford withdrew that favor, the result was instructive. Their models named the relevant condition in 94.9 percent of written scenarios. Ordinary people using the same models to work through the same scenarios got there less than 34.5 percent of the time, which was no better than a control group left to its own devices.16 Fairness requires mention of one study, from Google and a Harvard hospital, in which a chatbot interviewed a hundred real patients by text before their appointments, under live physician supervision, with respectable results. It had no comparison arm, and the physicians' plans were judged more practical and less costly.17 I do not suppose the authors of these papers to be propagandists. The effect of the headlines, whatever anyone intends, is to inform the public that its doctors are obsolete and to inform the doctors of the same.
Dr. Nundy, to his credit, saw an uninsured woman in a fifteen-minute slot, did not bury her in scans she could not afford, and sent her home better. He then drew the wrong lesson with great care. He recalls a professor who said that teaching a computer to play chess was easy and that the hard part was getting the robot to stroke its beard. In his telling the machine played the chess, meaning the diagnosis, while he supplied the beard, meaning the watching of faces and the earning of trust. He concedes that the diagnosis required knowledge and pattern recognition beyond his own, and he resolves to let the software play more of the chess while he remembers how to be a physician. Here, then, is the proposed division of labor: the software will practice medicine, and the doctor will furnish the facial expressions. The one piece of real medicine in his story is the one he prices lowest. He took the history, through an interpreter, and compressed it into the few tidy lines that made the question answerable, so that his chatbot, like the chatbots in the studies, was handed a vignette by a doctor. Knowing what is wrong with the patient was, until quite recently, the job. I can think of no interrogation technique, waterboarding included, that would extract from me a signed statement, submitted to the most prestigious journal in medicine, that a chatbot had to introduce me to the commonest cause of chronic stomach pain.
The author's other employer deserves a sentence. Humana announced in August that its new Chief Medical Officer would help shape the company's "use of AI."18 The company is the country's second-largest Medicare Advantage insurer, and it is defending a class action alleging that it used an algorithm to cut off elderly patients' rehabilitation coverage over the judgment of their physicians, an allegation it denies.1920 I hold no objection to a firm pursuing its interest, and the journal did print the affiliation. The reader should simply notice the convenience of the parable. A medicine in which judgment lives in software and the physician contributes warmth is a medicine in which judgment can be owned, standardized, and overruled by whoever holds the license. An insurer could hardly have commissioned a better fable, though it perhaps could have picked a more elusive diagnosis for the LLM to uncover.
A reader might fairly ask how a physician stumped by the commonest explanation for the commonest complaint comes to direct the clinical affairs of one of the nation's great insurers. The answer is that the two facts have nothing to do with each other, and that is the indictment. In a market where patients chose their physicians and paid them, a doctor's standing would rise and fall on what happens at the bedside. In the system we have built, the customers of a physician executive are boards, regulators, and editors, and the currency is titles, advisory panels, and essays. The release announcing his appointment lists counsel given to the Food and Drug Administration and the World Bank.21 Neither institution has ever had a stomachache.
Physicians will recognize the genre, because the fable arrives at the end of a long sequence. Over a generation the American doctor has lost his practice to the hospital and his hospital to the conglomerate. He works for a man with a business degree and a backpack who decides what he may order. He signs the charts, and carries the liability, of practitioners with a fraction of his training who were hired to replace him. Medicare has let his fees fall by a third against the cost of practice since 2001, while the patient's bill has traveled in the other direction and the public has somehow concluded that the doctor is the reason.22 When he objects, he is assigned a wellness module. Now he is informed, in the leading journal of his own profession and by an insurance executive, that the thinking can be left to the software and that his remaining contribution is a sympathetic face. I doubt that any committee planned this, and none was needed. A system in which third parties pay has little use for a physician's independent judgment and every use for his signature, and it will welcome any technology that promises the second without the first.
The journal believed it was publishing a parable about the limits of machines. It published one about the condition of the profession and the company it now keeps. Above the clinic sit executives who have drifted so far from the bedside that the commonplace reaches them as revelation. Beside them stand enthusiasts who never stood at a bedside and cannot tell the commonplace from discovery. A prestigious journal notarizes the exchange, and each party borrows from the other the credibility it lacks: the executive acquires the glamour of the technology, and the technologist acquires the imprimatur of medicine. A machine was asked what the textbook says, and it answered. The remarkable part is who had to ask, and who applauded. The patient, at least, is better. She was cured by an old pill and a doctor who finally said the obvious, and neither of those is a new technology.
Nundy S. What the Robot Still Can't Do. N Engl J Med. Published online November 14, 2026. doi:10.1056/NEJMp2606894. https://www.nejm.org/doi/full/10.1056/NEJMp2606894
See, for example, Topol E. Post on X. November 2026. https://x.com/erictopol/status/2121637287268389318; and Marchione M. Post on X. November 2026. https://x.com/maxmarchione/status/2121789282906189969
Longstreth GF. Functional Dyspepsia: Managing the Conundrum. N Engl J Med. 2006;354(8):791-793. doi:10.1056/NEJMp058326. https://www.nejm.org/doi/full/10.1056/NEJMp058326
Nundy S. What the Robot Still Can't Do. N Engl J Med. Published online November 14, 2026. doi:10.1056/NEJMp2606894. https://www.nejm.org/doi/full/10.1056/NEJMp2606894
Nundy S. What the Robot Still Can't Do. N Engl J Med. Published online November 14, 2026. doi:10.1056/NEJMp2606894. https://www.nejm.org/doi/full/10.1056/NEJMp2606894
Pasricha PJ, Talley NJ. Functional Dyspepsia. N Engl J Med. 2026;394(2):166-176. doi:10.1056/NEJMcp2501860. https://www.nejm.org/doi/full/10.1056/NEJMcp2501860
Espay AJ, Aybek S, Carson A, et al. Current Concepts in Diagnosis and Treatment of Functional Neurological Disorders. JAMA Neurol. 2018;75(9):1132-1141. https://pubmed.ncbi.nlm.nih.gov/29868890/
Hingray C, Popkirov S, Kozlowska K, et al. Functional/dissociative seizures: Proposal for a new diagnostic label and definition by the ILAE task force. Epilepsia. 2025. https://eprints.whiterose.ac.uk/id/eprint/231023/
Drossman DA. Functional Gastrointestinal Disorders: History, Pathophysiology, Clinical Features, and Rome IV. Gastroenterology. 2016;150(6):1262-1279. doi:10.1053/j.gastro.2016.02.032. https://pubmed.ncbi.nlm.nih.gov/27144617/
Volkers N. By any other name: Improving access to treatment and reducing stigma in psychogenic non-epileptic seizures. Epigraph. 2021;23(2). International League Against Epilepsy. https://www.ilae.org/journals/epigraph/epigraph-vol-23-issue-2-spring-2021/by-any-other-name-improving-access-to-treatment-and-reducing-stigma-in-psychogenic-non-epileptic-seizures
OpenAI. An OpenAI model has disproved a central conjecture in discrete geometry. May 20, 2026. https://openai.com/index/model-disproves-discrete-geometry-conjecture/
Alon N, Bloom TF, Gowers WT, et al. Remarks on the disproof of the unit distance conjecture. arXiv. Preprint posted online May 2026. arXiv:2605.20695. https://arxiv.org/abs/2605.20695
Hayek FA. The Use of Knowledge in Society. Am Econ Rev. 1945;35(4):519-530. https://www.jstor.org/stable/1809376
Goh E, Gallo R, Hom J, et al. Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial. JAMA Netw Open. 2024;7(10):e2440969. doi:10.1001/jamanetworkopen.2024.40969. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2825395
Microsoft AI. The Path to Medical Superintelligence. June 30, 2025. https://microsoft.ai/news/the-path-to-medical-superintelligence/
Bean AM, Payne RE, Parsons G, et al. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nat Med. 2026;32(2):609-615. doi:10.1038/s41591-025-04074-y. https://doi.org/10.1038/s41591-025-04074-y
Brodeur P, Koshy JM, Palepu A, et al. A prospective clinical feasibility study of a conversational diagnostic AI in an ambulatory primary care clinic. arXiv. Preprint posted online March 9, 2026. arXiv:2603.08448. https://arxiv.org/abs/2603.08448
Humana Inc. Humana Names Dr. Shantanu Nundy as Chief Medical Officer. Press release. August 20, 2026. https://policy.humana.com/news-and-resources/news-press/2026/humana-names-dr--shantanu-nundy-as-chief-medical
Humana used AI tool from UnitedHealth to deny Medicare Advantage claims, lawsuit alleges. Becker's Payer Issues. December 2023. https://www.beckerspayer.com/payer/humana-uses-ai-algorithm-from-unitedhealth-to-deny-medicare-advantage-claims-lawsuit-alleges/
Barrows et al v Humana Inc, No. 3:23-cv-00654 (WD Ky). Health Care Litigation Tracker, Georgetown University. https://litigationtracker.law.georgetown.edu/issues/artificial-intelligence-medicare-advantage
Humana Inc. Humana Names Dr. Shantanu Nundy as Chief Medical Officer. Press release. August 20, 2026. https://policy.humana.com/news-and-resources/news-press/2026/humana-names-dr--shantanu-nundy-as-chief-medical
American Medical Association. Medicare Physician Updates Compared to Inflation in Practice Costs (2001-2026). Updated January 2026. https://www.ama-assn.org/system/files/2026-medicare-updates-inflation-chart.pdf


