Integrating generative AI Into clinical reasoning education for medical students: Mixed methods study

“To account for multiple comparisons across the subscale pre-post tests, P values were adjusted using the Benjamini-Hochberg false discovery rate procedure [23], with Bonferroni correction reported for comparison; statistical significance was defined as P <.05.”

“With reference to pragmatic data-adequacy benchmarks for interview studies [22], we judged this corpus sufficient for the descriptive, workshop-specific aims of this phase, while acknowledging the modest number of interviews (see Limitations section).”

“The adequacy of the qualitative dataset was appraised using the concept of information power across the 5 dimensions described by Malterud et al [21].”

“From a cognitive load perspective, AI can reduce the extraneous load of information retrieval [29], but this convenience may also invite cognitive offloading; mandating verification positions AI as a scaffold for, rather than a substitute for, the development of clinical reasoning [30].”

“Reviews of AI in health care have attributed this momentum to rapid advances in data availability and natural language processing, which have enabled increasingly sophisticated support for knowledge-intensive clinical tasks [2].”

“Qualitative data from interviews and reflective narratives were analyzed using reflexive thematic analysis, following Braun and Clarke [24-26].”

“AI literacy was assessed as self-perceived competence, using a 20-item self-report questionnaire adapted from the Meta AI Literacy Scale [15] and informed by an established AI literacy framework [16]; hereafter, this outcome is referred to as self-perceived AI literacy.”

“The study adhered to the GRAMMS (Good Reporting of a Mixed Methods Study) guidelines [14] (Checklist 1) to ensure transparency and rigor in the integration of qualitative and quantitative data.”

“A central educational concern in AI integration is automation bias, the tendency to overrely on automated outputs without adequate scrutiny [27].”

“Reflective writings further suggested emerging professional identity adaptation [32], with students moving from sole knowledge holders toward “AI-augmented decision makers”; some voiced concern about overreliance (“we might lose our ability to research original textbooks”), while others described…”

“A substantial body of research on cognitive bias describes how anchoring, premature closure, and related heuristics can derail both modes of reasoning [10].”

“From a cognitive load perspective, AI can reduce the extraneous load of information retrieval [29], but this convenience may also invite cognitive offloading; mandating verification positions AI as a scaffold for, rather than a substitute for, the development of clinical reasoning [30].”

“Systematic comparisons show variable performance across clinical decision-support tasks [6], and randomized trials indicate that simply providing clinicians with access to LLMs does not consistently improve diagnostic accuracy compared with conventional information resources [5].”

“…the baseline floor effect and limited statistical power, this is most consistent with, though it cannot confirm, the premise that AI literacy and patient-centeredness need not be antagonistic competencies when instruction explicitly integrates ethical reflection and communication practice [33].”

“AI literacy was assessed as self-perceived competence, using a 20-item self-report questionnaire adapted from the Meta AI Literacy Scale [15] and informed by an established AI literacy framework [16]; hereafter, this outcome is referred to as self-perceived AI literacy.”

“Within medical education, clinical reasoning is conceptualized as a dynamic interplay between rapid pattern-recognition processes (System 1) and slower, analytical verification processes (System 2) [9].”

“Learners report limited competence in prompt design, bias appraisal, and ethical boundary setting when interacting with GenAI systems [11,12], and a recent multi-institutional study confirms that positive student attitudes coexist with persistent concerns and unmet training needs [13].”

“Because both analyses relied on the same participants, the CFA cannot independently confirm the factor structure and may capitalize on sample-specific characteristics [34].”

“Risks including hallucinations, embedded biases, and opaque reasoning may compromise patient safety if AI outputs are accepted uncritically [1,7,8].”

“Systematic comparisons show variable performance across clinical decision-support tasks [6], and randomized trials indicate that simply providing clinicians with access to LLMs does not consistently improve diagnostic accuracy compared with conventional information resources [5].”

“This pattern aligns with the concept of calibrated trust in human-AI systems, whereby appropriate reliance is dynamically adjusted to task context and risk [28].”

“Experimental diagnostic vignettes further demonstrate their potential: in controlled comparisons, GenAI systems have matched or outperformed physicians on structured case assessments [4,5].”

“This rapid adoption has prompted growing interest in how future physicians should be prepared to engage with AI-enabled clinical environments, and recent frameworks have begun to define the digital and AI competencies expected of medical graduates [3].”

“Learners report limited competence in prompt design, bias appraisal, and ethical boundary setting when interacting with GenAI systems [11,12], and a recent multi-institutional study confirms that positive student attitudes coexist with persistent concerns and unmet training needs [13].”

“Learners report limited competence in prompt design, bias appraisal, and ethical boundary setting when interacting with GenAI systems [11,12], and a recent multi-institutional study confirms that positive student attitudes coexist with persistent concerns and unmet training needs [13].”

“Risks including hallucinations, embedded biases, and opaque reasoning may compromise patient safety if AI outputs are accepted uncritically [1,7,8].”

“Risks including hallucinations, embedded biases, and opaque reasoning may compromise patient safety if AI outputs are accepted uncritically [1,7,8].”

“Collaborative learning attitudes were measured using 6 items adapted from the Collaborative Learning Attitude Scale (CLAS) [19,20], selected a priori based on item relevance; as with PPOS-D6, complete pre-post data were available for the same subset of participants (n=46).”

“Patient-centered attitudes were assessed using the 6-item Patient-Practitioner Orientation Scale-Short (PPOS-D6) [17, 18].”

“Collaborative learning attitudes were measured using 6 items adapted from the Collaborative Learning Attitude Scale (CLAS) [19,20], selected a priori based on item relevance; as with PPOS-D6, complete pre-post data were available for the same subset of participants (n=46).”

“Patient-centered attitudes were assessed using the 6-item Patient-Practitioner Orientation Scale-Short (PPOS-D6) [17, 18].”

“Recent work integrating verification-oriented GenAI activities within clinical reasoning education provides evidence for the feasibility of explicitly incorporating such practices into medical education (Liu et al., 2026).”