Teaching the Override Automation Bias, Clinical Reasoning, and the Formation of the Physician in the AI Era

In the summer of 2025, a single-blind randomized clinical trial published in NEJM AI delivered a finding that medical education cannot responsibly ignore. Forty-four physicians, all of whom had completed a 20-hour AI literacy training program, were randomly assigned to diagnose clinical vignettes using ChatGPT-4o alongside conventional resources. Those who received AI suggestions containing deliberately seeded errors exhibited significant diagnostic reasoning degradation compared to peers receiving error-free AI advice. Automation bias was not eliminated by AI literacy training (Qazi et al., 2025). This article, the fourth in a series on experiential intelligence in AI-mediated learning environments (Elliott, 2026a, 2026b, 2026c), argues that automation bias is now a core curricular concern for undergraduate medical education and proposes a structured developmental framework for teaching the override capacity, the willingness and ability to disagree with a confident wrong machine answer. Drawing on the Qazi et al. (2025) NEJM AI trial, the Goh et al. (2024) multisite JAMA Network Open randomized trial showing that GPT-4 alone outperformed physician-AI teams in diagnostic reasoning, the human factors literature on automation bias (Parasuraman & Manzey, 2010; Gaube et al., 2021; Küücking et al., 2024), the emerging concern about AI-related deskilling in medical training (Nguyen, 2024; Kalogeropoulos et al., 2025), and the Dreyfus model of skill acquisition (Dreyfus & Dreyfus, 1986; Benner, 1984), the article advances:(1) An empirical case that general AI literacy training, as currently practiced, does not eliminate automation bias in clinical reasoning. Override capacity must be trained directly as a distinct, domain-specific skill.(2) Seven training principles for building override capacity: build clinical reasoning first and AI integration second; train override on seeded errors rather than general critique; make the disagreement explicit; vary AI accuracy across training cases; train across cognitive load; provide structured feedback on override decisions; and assess override under unaided and AI-mediated conditions separately.(3) A four-stage developmental progression across undergraduate medical education: Foundation Without AI (preclinical years), Structured Comparison (transition to clinical work), Override Training (core clinical years), and Integrated AI-Mediated Practice (final year and into residency). The article concludes with implications for faculty development, program leadership, and accreditation, including the observation that a widening gap between unaided and AI-mediated learner performance is a curricular warning sign that AI is being used as a substitute for, rather than a check on, clinical reasoning. It is intended for medical educators, clerkship directors, curriculum committees, faculty developers, and program leadership in undergraduate and graduate medical education.