Keine Meisterschaft ohne die Maschine — Medizinische Ausbildung im Zeitalter der klinischen KI

NEJM AI 2026;3(11) DOI: 10.1056/AIp2600773 PERSPECTIVE No Mastery without the Machine — Medical Education in the Age of Clinical AI Philip Jarrett , M.D., M.B.A.,1 D. Mark Courtney , M.D., M.Sci.,1 Deborah Diercks , M.D., M.S., M.B.A.,1 and Samia Farooqi , M.D.1 Received: June 6, 2026; Revised: August 26, 2026; Accepted: September 9, 2026; Published: October 13, 2026 Abstract AI is often framed as a threat to medical education owing to concerns it may diminish the ­cognitive struggle through which trainees traditionally develop clinical judgment — a legiti- mate but incomplete concern. Medicine has never been defined by unaided human cognition; it has been defined by the obligation to use the best available tools in the service of patients. As AI systems increasingly extend not only clinicians’ perception, but also their reasoning, medical mastery may shift from memorization and independent synthesis toward the orches- tration, supervision, and calibrated deference to intelligent systems. The task for medical education is therefore not to preserve cognitive struggle for its own sake, but to refocus it from recalling facts to interrogating machine reasoning, detecting failure modes, and know- ing when human intervention improves care versus when it merely interrupts a better system. Introduction F or more than a century, medical education has treated cognitive struggle as a neces- sary element of professional formation. Students memorize biochemical pathways, defend differential diagnoses, present patients on rounds, and learn through the discomfort of being uncertain in front of others. This struggle is not incidental. It is one way trainees learn judgment, humility, and the limits of pattern recognition. The arrival of generative AI has therefore prompted understandable concern. If trainees can ask a model to generate a differential diagnosis, critique a management plan, summarize a record, or draft a note, will they still develop the cognitive muscle required to practice med- icine? Will a generation trained with AI become clinically fluent but intellectually hollow?1 These are serious concerns. But they risk mistaking the preservation of a training method for the preservation of medical mastery itself. The author affiliation is listed at the end of the article. Philip Jarrett can be contacted at AI and the Changing Nature of Clinical Expertise philip​.jarrett@utsouthwestern​.edu Medicine has never been defined by unaided human cognition, but by the obligation to or at Department of Emergency Medicine, UT Southwestern use the best available tools in the service of patients. The stethoscope extended hearing. Medical Center, 5323 Harry Hines The electrocardiogram extended perception. The laboratory extended inference. Imaging Blvd E4.300, Dallas, TX 75390. NEJM AI is produced by NEJM Group, a division of the Massachusetts Medical Society. Downloaded from ai.nejm.org by Christian Elsner on October 26, 2026. extended anatomy beyond the visible body. The electronic The task, then, is not to remove struggle from medical edu- health record, despite its many failures, extended memory. cation. It is to refocus it. Future trainees may need less fric- Each technology displaced some prior form of clinical skill tion around memorizing facts that machines can retrieve while also giving clinicians new skills to hone. instantly and more friction around interrogating machine reasoning, reconciling contradictory outputs, identifying Artificial intelligence may be different in degree, but not hidden assumptions, detecting automation bias, and decid- in kind. It does not merely extend the physician’s senses; it ing when deference to AI is safer than human correction. increasingly extends the physician’s reasoning. That is pre- The educational objective should not be unaided cognition cisely why it feels threatening. AI enters the territory phy- for its own sake, but calibrated judgment in the presence of sicians have long regarded as most central to professional powerful cognitive machinery. identity: the capacity to synthesize incomplete informa- tion, generate hypotheses, estimate risk, and recommend action.2 Yet the relevant question is not whether AI will erode Implications for Medical Education some traditional forms of expertise. It will. The question is and Practice whether medical education will defend yesterday’s exper- This shift also exposes the limitations of current assess- tise as a marker of professional legitimacy or cultivate the ment and supervision. A curriculum organized around expertise required for tomorrow’s care. closed-​book recall and standardized testing may be poorly The analogy may be less chess than aviation. We do not suited to an era in which the central competency is not train pilots to navigate primarily by stars because celestial memory alone, but combined human and machine judg- navigation once represented mastery. We train them to ment. Training will require faculty who can model nonde- understand aircraft systems, weather, automation, com- fensive AI use at the bedside and beyond, AI systems that munication, failure modes, and the conditions under which serve as tutors as well as tools, and assessments that test human intervention is lifesaving or catastrophic. The mod- whether physicians can recognize unsupported recommen- ern pilot is not less skilled because autopilot exists. The dations, explain algorithmic reasoning, and justify accept- locus of skill has shifted from continuous manual control to ing or overriding a machine-​generated plan. systems command. The evidence base is still incomplete, but the trajectory is Medicine may be entering a similar transition. The phy- difficult to ignore. Recent studies suggest that frontier lan- sician of the near future may not be the person who inde- guage models can match or exceed physicians on some pendently recalls the longest differential diagnosis or complex diagnostic and management reasoning tasks.4 personally integrates every clinical variable. Instead, the Other work has shown that physician use of AI is not auto- physician may be the person who can orchestrate intelligent matically beneficial; the human–AI team can perform worse systems: directing AI agents to obtain history, interrogate when clinicians misunderstand, underuse, or overtrust the medical record, identify inconsistencies, generate diag- model output.5 In the field of surgery, AI-​enabled robotic nostic and therapeutic pathways, surface uncertainty, com- systems have begun performing complex procedural tasks municate trade-​offs, and monitor for deterioration (Fig. 1). in controlled settings with increasing autonomy.6 None of This is not “never skilling.”3 It is a different skill. this proves that AI is ready to replace clinicians in ordinary practice. Current systems remain brittle, variably evalu- ated, unevenly integrated into workflows, and vulnerable to failure modes that may be invisible to patients and clini- Refocusing the Struggle cians alike. This reframing should not become an excuse for friction- But dismissing these systems because they are not yet com- less training. A trainee who has never struggled to form plete is like dismissing the Wright Flyer because it could not an argument, defend a differential diagnosis, or sit with cross the Atlantic. The relevant fact is not that early aircraft uncertainty may be poorly prepared to recognize when were fragile; it is that flight had become possible. an AI system is incomplete, overconfident, or wrong. The danger is not that trainees will use AI. The danger is that The most comforting arguments against medical AI are they may never learn what it feels like to reason before AI also the most precarious. We say AI cannot understand the intervenes. physician–patient relationship. We say it cannot do what NEJM AI 2 NEJM AI is produced by NEJM Group, a division of the Massachusetts Medical Society. Downloaded from ai.nejm.org by Christian Elsner on October 26, 2026. Figure 1. Physician Judgment in an AI-​Mediated Clinical Encounter. This figure shows a hypothetical clinical interface in which AI generates diagnostic and management recommendations while the physician evaluates their appropriateness, incorporates patient priorities, and judges when to defer or intervene. Displayed findings and confidence estimates are illustrative. EHR denotes electronic health record; and SGLT2, sodium–glucose cotransporter-​2. we do. We say patients need human presence, moral judg- This possibility should not be caricatured as technologi- ment, and empathy. In many clinical contexts, these claims cal surrender. It is a call for professional humility. Medical remain true. But they can also become professional defense education should not teach students to compete with AI on mechanisms — ways of converting current indispensability tasks machines may soon perform better. It should teach into a permanent moral principle. them to evaluate, supervise, challenge, integrate, and, when necessary, yield to AI. The physician–patient relationship is sacred because it serves the patient, not because it preserves the physician. Yielding to AI is undoubtedly an uncomfortable dynamic If an AI system someday diagnoses more accurately, treats to accept. Physicians are trained to act, decide, rescue, and more safely, monitors more continuously, communicates take responsibility. We are less often trained to recognize more patiently, and remembers more completely than a when our intervention adds noise to a better system. Yet human clinician, our ethical obligation will not be to protect the future of medicine may require a new form of restraint: the primacy of the physician; it will be to ask what role the knowing when human judgment improves care and when it physician should continue to occupy. merely interrupts it. NEJM AI 3 NEJM AI is produced by NEJM Group, a division of the Massachusetts Medical Society. Downloaded from ai.nejm.org by Christian Elsner on October 26, 2026. References The Future Role of the Physician 1. Keren R, Desai BR, West DC. Promoting clinical expertise in the age of AI: no struggle, no mastery. JAMA 2026;335:1837-​1838. For centuries, physicians have stood at the center of med- DOI: 10​.1001​/jama​.2026​.6097. icine’s cognitive architecture. AI asks whether that posi- 2. Brodeur PG, Buckley TA, Kanjee Z, et al. Performance of a large tion is a law of nature or a historical phase.2 The profession language model on the reasoning tasks of a physician. Science can respond by insisting that authentic medical practice 2026;392:524-​527. DOI: 10​.1126​/science​.adz4433. requires unaided human mastery. Or it can prepare trainees 3. Ke Y, Jin L, Chiat Ling Ong J, et al. AI-induced never-skilling in for a world in which mastery means knowing how to com- medical education. Nat Med 2026;32:1997–2006. DOI: 10.1038/ mand systems more capable than oneself. s41591-026-04438-y. The first path protects the mythology of the physician. The 4. Goh E, Gallo R, Hom J, et al. Large language model influence second protects patients. on diagnostic reasoning: a randomized clinical trial. JAMA Netw Open 2024;7:e2440969. DOI: 10​.1001​/jamanetworkopen​.2024​ One day, the most dangerous actor in the room may not .40969. be the machine making a recommendation; it may be the human who insists on overruling it simply because being in 5. Rodman A, Beam AL, Manrai AK. Exploring the past — and future — of medical diagnosis and artificial intelligence. NEJM AI the loop still feels like being in charge. 2024;1(9). DOI: 10​.1056​/AIp2400707. 6. Kim JWB, Chen JT, Hansen P, et al. SRT-​H: a hierarchical frame- Disclosures work for autonomous surgery via language-​conditioned imitation Author disclosures are available at ai​.nejm​.org. learning. Sci Robot 2025;10:eadt5254. DOI: 10​.1126​/scirobotics​ .adt5254. ChatGPT Images 2.5 (OpenAI) was used to assist with image genera- tion. No AI tools were used for writing assistance. Author Affiliation 1 Department of Emergency Medicine, The University of Texas Southwestern Medical Center, Dallas, USA NEJM AI 4 NEJM AI is produced by NEJM Group, a division of the Massachusetts Medical Society. Downloaded from ai.nejm.org by Christian Elsner on October 26, 2026.