The cognitive ecology of medicine: generative AI, clinical reasoning, and the preservation of adaptive expertise

Abstract Generative AI (genAI) does not simply add another layer of decision support to medicine; it changes the environment in which clinical reasoning is performed, taught, and judged. This matters because clinical reasoning is not simply the production of answers, but the management of uncertainty in a particular patient: framing the problem, deciding what matters most, distinguishing between plausible alternatives, calibrating uncertainty, and revising plans as new information emerges. As large language models (LLMs) become embedded in clinical workflow, the work of reasoning is not removed but redistributed. Clinicians may do less first-pass synthesis but more verification, contextualization and judgment of machine-generated outputs. This may reduce burden, but it also increases vulnerability to anchoring, automation bias, and premature closure. The educational risk is therefore not only that AI may be wrong, but that routine answer-first use may weaken the developmental work through which adaptive expertise forms. In training, this may lead to deskilling, never-skilling, or mis-skilling unless key domains of reasoning remain visible under supervision. We argue that these include framing, weighting salient findings, discriminating between alternatives, calibrating uncertainty, and revising plans in context. In this light, one promising educational use of LLMs may be as question-prompting aids used after trainee commitment to test and deepen reasoning. The future of clinical reasoning in the age of genAI will depend less on what the machine can produce than on whether medical education preserves the conditions under which human judgment is formed.