The emergence of artificial intelligence (AI), particularly generative AI (GenAI), represents an inflection point for medical education. AI is reshaping how trainees acquire knowledge, synthesize information, and participate in clinical decision-making. In order to master this shift in learner experience, curricular response must move beyond passive accommodation.
In simple terms, AI can be thought of as a cognitive multiplier. Contemporary systems can summarize literature, generate differential diagnoses, and otherwise provide infinite riffs on metaphors or learning styles to connect with learners in ways we can’t. AI also approximates one-on-one tutoring, adapts to learner pace, and provides iterative feedback—capabilities that traditional curricula, constrained by faculty time and clinical workload, all struggle to deliver.
AI capabilities, however, introduce a paradox. The efficiency of AI-mediated cognition risks attenuating the same processes that medical education is designed to cultivate. Multiple papers have raised concern that reliance on AI outputs can erode critical thinking and independent knowledge construction; these erosions are deeper in early learners.1 Put practically, diagnostic error carries downstream patient harm; however, if there isn’t a proper base, the learner doesn’t know the error has happened until it is too late in the Bayesian update. Erosion of skills isn’t abstract; it’s a quality and safety issue.
AI also has wide and severe effects on assessment. Traditional assessments, including written assignments and remote multiple-choice exams, are increasingly vulnerable to AI augmentation or override. There is now a small but not insignificant shift to assessments that are impervious to assault by AI, including viva voce, blue books, and rigorous, observed, clinical encounters.2
A third domain is faculty development. Data suggest that adoption of AI among faculty remains uneven; there is still a substantial proportion of educators lacking familiarity or direct experience with these tools.3 This creates misalignment between learner behavior and faculty oversight. Trainees are already using AI overwhelmingly. Combine that with the above, and you get learners using it informally and without guidance; in many cases, we don’t have the skills to even broadly supervise its use. Curricular reform, therefore, must include parallel investment in faculty development. Without this, institutions risk creating another arm of the hidden curriculum, where AI use is both ubiquitous and largely unexamined.
Ethics and professionalism represent the fourth domain (there may be more). AI introduces new risk vectors, including hallucinated information, embedded bias, data privacy concerns, and ambiguity in accountability, all of which are potential risks to the ethical practice of medicine.4
AI is also surprisingly good at “pretending,” or choosing words in a probabilistic manner, to seem ethical. The palliative care team here at Columbia has trialed AI to see what it will generate in incredibly difficult circumstances; the results were interesting. For example, a patient at the end of life kept saying they were a fighter, they had been a fighter their whole life, and against the counsel of everyone, they wanted every possible treatment. As a test, the scenario was fed into a commercial large language model, and the response “It might be important now to ask yourself what being a fighter means” is considered a question seasoned palliative care attendings would ask.
These issues are inseparable from professional identity formation. The use of AI in clinical care raises questions of authorship, responsibility, and trust. Who is accountable for a decision? How should uncertainty be communicated to patients? How should a potential diagnosis and treatment plan be credited? All of these questions are central to the formation of the eventual physician.
It’s not all bad. AI does create opportunities to enhance assessment itself. Generative models can construct clinical cases, simulate patient interactions, and provide structured feedback on learner performance. Early work in graduate medical education suggests that AI-enabled simulation can expand access to deliberate practice and, as importantly, standardize feedback.5 This can be particularly relevant in hospital medicine, where variability in patient exposure can limit experiential learning. AI may allow programs to “fill gaps” in clinical experience, ensuring exposure to high-value conditions and decision points.
From a systems perspective, AI intersects with all of our workflows. Opportunities include reduction of documentation burden, enhanced decision support, and real-time analytics. These changes will inevitably alter the cognitive environment in which trainees learn. If clinical reasoning is increasingly scaffolded by AI, the question becomes: what is the irreducible skill set of the physician? Curriculum must answer this explicitly. If not, attrition is assured, and entropy will do its damage.
In operational terms, several curricular strategies emerge. First, AI must not be used procedurally, but rather reflectively. It is useful and appropriate for AI to be a friendly adversary. “Prove my thinking wrong right now.” “Are there any cognitive or other biases my thinking suggests?” After the output is generated, we encourage learners to ask, “Do I trust this? Why or why not?” If the answer was right, did the learner fight against it at any point? If so, why? If the answer was wrong, we teach the learners that it is usually plausibly wrong. This plausibility can be instructive. This overall metacognition allows us to turn a ChatGPT session into professional identity formation.
Since AI is not a neutral tool, it must be explicitly taught as an object of study. At Columbia, we teach our learners to ask the same two questions about AI that the Amish use when a new tool is introduced into their culture: 1) What is the purpose of this tool? 2) How does this tool change me as I use it? By having the learners and the faculty continually ask themselves these questions, we aim to minimize deskilling while at the same time encouraging metacognition so the user can engage with AI as a force multiplier.
In summation, AI is not simply another educational technology. It is a reconfiguration of the cognitive infrastructure of medicine. The curricular response must be equally structured. For hospitalists—who already operate at the nexus of care delivery and education—this represents both a challenge and an opportunity. The goal is not to produce physicians who can use AI, but physicians who can think in an AI-mediated environment.
Dr. Migliore is an assistant professor of medicine at Columbia University College of Physicians and Surgeons and medical director of New York Presbyterian Intelligence at Columbia University Medical Center–New York Presbyterian. He is also chair of AI for Learning (undergraduate medical education) at Vagelos College of Physicians and Surgeons, Columbia University.
References
- Naqvi WM, et al. Critical thinking in the age of generative AI: implications for health sciences education. Front Artif Intell. 2025;8:1571527. doi:10.3389/frai.2025.1571527.
- Brumby D, et al. Sampled vivas are pivotal in combating AI cheating. Times Higher Education website. https://www.timeshighereducation.com/opinion/sampled-vivas-are-pivotal-combating-ai-cheating. Published March 6, 2025. Updated March 6, 2025. Accessed March 29, 2026.
- Mulford D. AI in higher education: a meta summary of recent surveys of students and faculty. Campbell University website. https://sites.campbell.edu/academictechnology/2025/03/06/ai-in-higher-education-a-summary-of-recent-surveys-of-students-and-faculty/. Published March 6, 2025. Accessed March 29, 2026.
- Mohsin Khan M, et al. Towards secure and trusted AI in healthcare: A systematic review of emerging innovations and ethical challenges. Int J Med Inform. 2025;195:105780. doi:10.1016/j.ijmedinf.2024.105780.
- Hicke Y, et al. MedSimAI: simulation and formative feedback generation to enhance deliberate practice in medical education. Arxiv.org website. doi.org/10.48550/arXiv.2503.05793. Published March 1, 2025. Updated January 22, 2026. Accessed March 29, 2026.