From compliance to competence: a multi-level model for AI governance in undergraduate medical education Authors: Juan José Gualda-Gea, Lourdes Estefanía Barón-Miras, Maria Jesús Bertran, Alberto L. Garcia-Basteiro, Guillermo Mena, Marta Aldea, Gemina Santana, Jaume Grau, Andrés Prat Marin, Isabel Torá-Rocamora, Antoni Trilla Published: 2026 ABSTRACT AI's fast-paced irruption in medicine and education has challenged undergraduate medical colleges, with institutions documenting policies emphasizing immediate conduct restrictions, while strategic oversight, infrastructure and evaluation remain less consistently developed. We propose a multi-level framework for AI governance in undergraduate medical education based on accountability, competence and co-responsibility across institutions, educators and students, designed to turn a fast-evolving technology from a threat into an opportunity for collective progress. From reaction to proaction in AI governance Artificial intelligence (AI) has already entered classrooms, clinical settings and examination halls1. Generative tools are now capable of drafting patient cases, simulating diagnostic reasoning and providing personalised training2. Yet the governance of these technologies in the university remains largely reactive, designed to prevent misconduct rather than to build capacity3. Recent evidence from US medical schools and Swedish higher education institutions suggests that institutional responses remain limited and uneven, often focusing on academic integrity while strategic oversight, infrastructure and implementation receive less attention4,5. Medicine has been impacted by disruptive technologies before, and its educational reaction has frequently been uneven. History offers several examples of tools that reshaped not only what physicians could do, but which skills they retained. The internet and, later, electronic health records broadened the availability, access and quality of health information. Yet, digitalization has made us dependable of computers, networks and electricity. Documentation is a leading driver of healthcare workers' burnout, being a cautionary example of technology imposed on clinicians rather than designed along with them6,7. Besides, digitalization has been unevenly implemented across territories and across countries, widening even more health inequity gaps8. Another example can be medical imaging, which revolutionised diagnosis, yet overreliance on it has been repeatedly linked to the loss of physical examination and bedside skills9. Imaging is increasingly ordered before, or instead of, clinical reasoning. How many times has a scan been requested “to see what it shows” instead of testing a preconceived hypothesis? This is precisely the pattern that AI threatens to boost, and precisely, the reason a governance framework must differentiate between competencies that could safely benefit from AI assistance and those that must remain AI-independent10,11. The challenge extends beyond academic integrity. As AI becomes embedded in healthcare, physicians must not only know how to use it but also when and whether to trust it, navigating the risks of automation bias and algorithmic opacity12. Without structured governance, universities risk producing doctors used to using these tools but unable either to harness their full potential or to recognize when they are being misled by them13. Governance should therefore evolve from reactive control mechanisms into a proactive process of education, reflection, and shared responsibility, aligned with global ethical principles and regulatory requirements14. The missing layer of governance Existing ethical and regulatory frameworks define safeguards such as privacy, fairness, safety and human oversight14,15 while professional and educational competency frameworks specify what clinicians, educators and learners should know and be able to do12. Within universities, governance remains fragmented and inconsistent. Evidence from U.S. medical schools suggests that few institutions maintain dedicated oversight structures for educational AI, while regular audits of fairness, reliability, and transparency remain uncommon4,16. Where directives exist, they often seek to minimize risk rather than cultivate capability. This reactive approach leaves educators and students in a moral “grey zone”, uncertain where legitimate learning assistance ends, and academic dishonesty begins3,17,18. Anchoring governance through accountability, competence and co-responsibility Hereby, we propose a framework structured around three interconnected core implementation principles: Accountability Trust but verify. Every AI tool integrated into medical education must face systematic scrutiny. Does it achieve its stated educational objectives? Does it introduce hidden harms, such as algorithmic bias, skill atrophy, or false confidence in flawed outputs? Accountability answers through regular audits, transparent performance metrics, and honest reckoning when tools fail. Operationally, this requires a body that assesses the proposed tools and the implementation of AI tool proposals, similar to classical biomedical Health Technology Assessment Committees. On the other hand, project leaders should define the intended use, review evidence, identify advantages, relevance, and benefits and, if accepted, monitor outcomes. Potential ethical risks must be listed, and all parts of the project must commit to preserve ethical integrity. Competence AI must empower, not diminish. Critical clinical judgement remains the irreducible foundation of medical education. Algorithmic tools should handle routine tasks to create space for higher-order clinical reasoning, strengthening rather than replacing essential skills. Judgement should be sharpened rather than autonomy eroded. Competence should therefore be demonstrated through observable performance of a defined task under stated conditions and against predefined criteria, rather than inferred solely from course attendance, learner confidence, or the quality of an AI-assisted product. Co-responsibility A pact, not a prosecution. AI governance in medical education requires shared commitment rather than top-down control. Universities provide infrastructure and guidance; educators model judicious integration; students engage tools with informed intentionality. Success emerges from collective ownership. Responsibilities must be specified along the different phases of implementation and follow-up, for both staff and users: approval, supervision, disclosure, verification and escalation. The institution retains final responsibility for the systems and the assessment conditions it authorises. A multi-level framework for AI governance Effective governance operates across three mutually reinforcing levels, institutional, curricular, and individual, guided by the principles of accountability, competence, and co-responsibility (Fig. 1).Fig. 1Multi-level framework for universities' AI governance in undergraduate medical education.Three foundational principles (accountability, competence and co-responsibility) frame the model and apply across every level of implementation. Within this frame, governance operates at three mutually reinforcing levels: institutional (protected time for educator AI literacy, equitable resource provision, entrustment-based tool validation and ethical oversight of academic monitoring), curricular (competency review, assessment redesign and the five fronts of that review, rethinking evaluation methods) and individual (professional responsibility, transparent engagement, and the protection of initiative and creativity). Bidirectional arrows denote reciprocal influence between levels: institutional policy shapes curricular and individual practice, while classroom and individual experience inform institutional decisions. The framework converges on its intended outcome: physicians who master AI, not mastered by it. Infrastructure and cybersecurity, interdisciplinary collaboration, and the challenges to implementation operate transversally across all levels. Figure created by the authors using Canva. Multi-level framework for universities' AI governance in undergraduate medical education. Three foundational principles (accountability, competence and co-responsibility) frame the model and apply across every level of implementation. Within this frame, governance operates at three mutually reinforcing levels: institutional (protected time for educator AI literacy, equitable resource provision, entrustment-based tool validation and ethical oversight of academic monitoring), curricular (competency review, assessment redesign and the five fronts of that review, rethinking evaluation methods) and individual (professional responsibility, transparent engagement, and the protection of initiative and creativity). Bidirectional arrows denote reciprocal influence between levels: institutional policy shapes curricular and individual practice, while classroom and individual experience inform institutional decisions. The framework converges on its intended outcome: physicians who master AI, not mastered by it. Infrastructure and cybersecurity, interdisciplinary collaboration, and the challenges to implementation operate transversally across all levels. Figure created by the authors using Canva. Operational tool implementation algorithm To operationalise the conceptual framework for specific AI uses, we propose a risk-proportionate decision algorithm embedded within broader technical and non-technical governance mechanisms (Fig. 2). The proponent first defines the intended use, after which the institutional AI governance body assigns a risk tier and specifies the minimum evidence and safeguards required. A proposed use level (sandbox-only, supervised/restricted, or routine use) is then defined together with its scope, required human oversight, educational outcomes, stop criteria, and accountable decision owner. Before proceeding, the proposed use must satisfy a non-negotiable floor encompassing legality, privacy, safety, non-discrimination, academic integrity, and meaningful human review. If this floor is not met, the use should be modified and reassessed where feasible or discarded when remediation is not possible. Uses that meet the floor proceed to a controlled pilot with predefined scope, duration, monitoring, and stop criteria. Pilot outcomes and safeguards are subsequently reviewed; unsuccessful uses may again be modified and reassessed or discarded. Successful uses proceed to implementation, documentation, monitoring, and periodic reassessment under a revocable institutional AI-entrustment decision.Fig. 2Risk-proportionate decision pathway for AI governance in undergraduate medical education.A proposed AI use progresses from definition of its intended purpose through risk classification, specification of minimum evidence and safeguards, and definition of a proposed use level and its conditions. Before a controlled pilot, the proposed use must satisfy a non-negotiable floor encompassing legality, privacy, safety, non-discrimination, academic integrity, and meaningful human review; uses that fail may be modified and reassessed or discarded when remediation is not feasible. Pilot outcomes and safeguards are subsequently evaluated using the same iterative logic. Successful uses proceed to implementation, documentation, monitoring, and periodic reassessment under a revocable institutional AI-entrustment decision. The pathway distinguishes the responsibilities of the proponent or requesting team (P) from those of the institutional AI governance body (G); coloured P and G markers identify the lead actor at each step, while paired P/G markers indicate shared responsibility. The callouts illustrate how potential harm should be balanced against potential benefit and how innovation should be reconciled with learner independence. Figure created by the authors using Microsoft PowerPoint. Risk-proportionate decision pathway for AI governance in undergraduate medical education. A proposed AI use progresses from definition of its intended purpose through risk classification, specification of minimum evidence and safeguards, and definition of a proposed use level and its conditions. Before a controlled pilot, the proposed use must satisfy a non-negotiable floor encompassing legality, privacy, safety, non-discrimination, academic integrity, and meaningful human review; uses that fail may be modified and reassessed or discarded when remediation is not feasible. Pilot outcomes and safeguards are subsequently evaluated using the same iterative logic. Successful uses proceed to implementation, documentation, monitoring, and periodic reassessment under a revocable institutional AI-entrustment decision. The pathway distinguishes the responsibilities of the proponent or requesting team (P) from those of the institutional AI governance body (G); coloured P and G markers identify the lead actor at each step, while paired P/G markers indicate shared responsibility. The callouts illustrate how potential harm should be balanced against potential benefit and how innovation should be reconciled with learner independence. Figure created by the authors using Microsoft PowerPoint. The pathway also makes trade-offs explicit. Where potential educational benefit conflicts with potential harm, responses may include data-minimised deployment, modification of the intended use, an approved alternative, or non-use. Where AI-supported learning or assessment threatens learner independence, curricula should identify which capabilities must remain independently demonstrable and assess AI-free performance separately from the ability to collaborate safely with AI. Figure 2 therefore translates the multilevel framework into an explicit, reviewable pathway for proposing, evaluating, implementing, and reassessing specific AI uses. At the institutional level, universities must establish: Protected time for AI literacy acquisition for teachers: protecting time for acquiring AI literacy for teachers is essential, not optional19. Most medical educators balance full-time professional practice with teaching20, Institutional policies assuming educators will develop AI competence without protected hours will produce underprepared educators and inadequate student outcomes.Equitable resources provision and funding: AI integration risks exacerbating educational inequities for students. Institutions must ensure equitable access, providing institutional licensing and structured training to prevent widening opportunity gaps.Extensively, clinical educators with partial academic dedication cannot reasonably fund AI tools while managing overburdened schedules21. Junior educators allocate most of their time to clinical practice20 while earning lower compensation, limiting capacity for self-funding. They depend on informal support networks, widening the opportunity breach.Paradoxically, early-career educators represent the highest-return investment for institutional AI resource provision in the long run, amplifying downstream impact on successive student cohorts. Strategic implementation and AI ‘entrustment’ AI entrustment is a revocable institutional authorisation for a specified AI system to support a defined educational task, population and level of supervision. It does not attribute agency to the technology; the permitted level of institutional reliance and the evidentiary requirements should remain proportionate to uncertainty, educational stakes, data sensitivity and irreversibility. Legitimate scepticism should be addressed through clear safeguards, defined use boundaries, and proportionate oversight. Universities should adapt entrustment principles to AI, authorising only the scope of use justified by evidence of the system’s performance, reliability and alignment with the intended educational purpose22. Independent pre-deployment evaluation should be followed by continuous monitoring, with students and educators contributing to validation, while institutions retain ultimate accountability. Ethical governance of academic monitoring Algorithmic assessment of student performance or potential can entrench bias, enable premature labelling, and undermine autonomy if deployed without safeguards14,15. Institutions should define a legitimate educational purpose, minimize longitudinal data collection, disclose how data and models are used, audit performance and subgroup effects, and provide routes for review and appeal. Final academic decisions must remain under accountable human authority. These safeguards should combine duty-based constraints with evaluation of consequences for learning, equity, learner agency and future patient care. Bias may arise from training data, model design, assessment labels or unequal access, requiring subgroup monitoring, human review and reassessment as evidence and technologies evolve14,15. Protected time for AI literacy acquisition for teachers: protecting time for acquiring AI literacy for teachers is essential, not optional19. Most medical educators balance full-time professional practice with teaching20, Institutional policies assuming educators will develop AI competence without protected hours will produce underprepared educators and inadequate student outcomes. Protected time for AI literacy acquisition for teachers: protecting time for acquiring AI literacy for teachers is essential, not optional19. Most medical educators balance full-time professional practice with teaching20, Institutional policies assuming educators will develop AI competence without protected hours will produce underprepared educators and inadequate student outcomes. Equitable resources provision and funding: AI integration risks exacerbating educational inequities for students. Institutions must ensure equitable access, providing institutional licensing and structured training to prevent widening opportunity gaps.Extensively, clinical educators with partial academic dedication cannot reasonably fund AI tools while managing overburdened schedules21. Junior educators allocate most of their time to clinical practice20 while earning lower compensation, limiting capacity for self-funding. They depend on informal support networks, widening the opportunity breach.Paradoxically, early-career educators represent the highest-return investment for institutional AI resource provision in the long run, amplifying downstream impact on successive student cohorts. Equitable resources provision and funding: AI integration risks exacerbating educational inequities for students. Institutions must ensure equitable access, providing institutional licensing and structured training to prevent widening opportunity gaps. Extensively, clinical educators with partial academic dedication cannot reasonably fund AI tools while managing overburdened schedules21. Junior educators allocate most of their time to clinical practice20 while earning lower compensation, limiting capacity for self-funding. They depend on informal support networks, widening the opportunity breach. Paradoxically, early-career educators represent the highest-return investment for institutional AI resource provision in the long run, amplifying downstream impact on successive student cohorts. Strategic implementation and AI ‘entrustment’ AI entrustment is a revocable institutional authorisation for a specified AI system to support a defined educational task, population and level of supervision. It does not attribute agency to the technology; the permitted level of institutional reliance and the evidentiary requirements should remain proportionate to uncertainty, educational stakes, data sensitivity and irreversibility. Legitimate scepticism should be addressed through clear safeguards, defined use boundaries, and proportionate oversight. Universities should adapt entrustment principles to AI, authorising only the scope of use justified by evidence of the system’s performance, reliability and alignment with the intended educational purpose22. Independent pre-deployment evaluation should be followed by continuous monitoring, with students and educators contributing to validation, while institutions retain ultimate accountability. Strategic implementation and AI ‘entrustment’ AI entrustment is a revocable institutional authorisation for a specified AI system to support a defined educational task, population and level of supervision. It does not attribute agency to the technology; the permitted level of institutional reliance and the evidentiary requirements should remain proportionate to uncertainty, educational stakes, data sensitivity and irreversibility. Legitimate scepticism should be addressed through clear safeguards, defined use boundaries, and proportionate oversight. Universities should adapt entrustment principles to AI, authorising only the scope of use justified by evidence of the system’s performance, reliability and alignment with the intended educational purpose22. Independent pre-deployment evaluation should be followed by continuous monitoring, with students and educators contributing to validation, while institutions retain ultimate accountability. Ethical governance of academic monitoring Algorithmic assessment of student performance or potential can entrench bias, enable premature labelling, and undermine autonomy if deployed without safeguards14,15. Institutions should define a legitimate educational purpose, minimize longitudinal data collection, disclose how data and models are used, audit performance and subgroup effects, and provide routes for review and appeal. Final academic decisions must remain under accountable human authority. These safeguards should combine duty-based constraints with evaluation of consequences for learning, equity, learner agency and future patient care. Bias may arise from training data, model design, assessment labels or unequal access, requiring subgroup monitoring, human review and reassessment as evidence and technologies evolve14,15. Ethical governance of academic monitoring Algorithmic assessment of student performance or potential can entrench bias, enable premature labelling, and undermine autonomy if deployed without safeguards14,15. Institutions should define a legitimate educational purpose, minimize longitudinal data collection, disclose how data and models are used, audit performance and subgroup effects, and provide routes for review and appeal. Final academic decisions must remain under accountable human authority. These safeguards should combine duty-based constraints with evaluation of consequences for learning, equity, learner agency and future patient care. Bias may arise from training data, model design, assessment labels or unequal access, requiring subgroup monitoring, human review and reassessment as evidence and technologies evolve14,15. At the curricular level: AI will reshape the competencies and content of the traditional undergraduate medical curricula. Consequently, evaluation methods will have to be reimagined:Competency review: Artificial intelligence will fundamentally reshape physician competencies. Not merely technical capabilities, but diagnostic reasoning, patient communication, and professional judgement. Medical education must evolve accordingly. This transformation resembles a “house moving“ rather than simple curricular addition: existing competencies must be rigorously evaluated, discarding those obsolete by algorithmic advancement; preserving those whose fundamental importance transcends technological change; and deliberately integrating new competencies essential for AI-enhanced healthcare delivery. E.g., which diagnostic skills retain irreplaceable value when algorithms can analyse imaging with superhuman accuracy? Which forms of clinical reasoning must we preserve as AI transforms information access? These questions demand ongoing attention. This process must remain organic and adaptive, as competency requirements will evolve continuously rather than stabilize. The integration of AI into medical education is inextricably linked to its implementation in clinical practice. Healthcare AI adoption will likely follow a non-linear path, constrained by data availability, workforce readiness, and infrastructure limitations that create friction between rapid technological development and measured clinical deployment. Medical curricula must therefore prepare students for an AI-integrated future while acknowledging the uneven, incremental reality of healthcare transformation. Competency review: Artificial intelligence will fundamentally reshape physician competencies. Not merely technical capabilities, but diagnostic reasoning, patient communication, and professional judgement. Medical education must evolve accordingly. This transformation resembles a “house moving“ rather than simple curricular addition: existing competencies must be rigorously evaluated, discarding those obsolete by algorithmic advancement; preserving those whose fundamental importance transcends technological change; and deliberately integrating new competencies essential for AI-enhanced healthcare delivery. E.g., which diagnostic skills retain irreplaceable value when algorithms can analyse imaging with superhuman accuracy? Which forms of clinical reasoning must we preserve as AI transforms information access? These questions demand ongoing attention. This process must remain organic and adaptive, as competency requirements will evolve continuously rather than stabilize. The integration of AI into medical education is inextricably linked to its implementation in clinical practice. Healthcare AI adoption will likely follow a non-linear path, constrained by data availability, workforce readiness, and infrastructure limitations that create friction between rapid technological development and measured clinical deployment. Medical curricula must therefore prepare students for an AI-integrated future while acknowledging the uneven, incremental reality of healthcare transformation. Competency review: Artificial intelligence will fundamentally reshape physician competencies. Not merely technical capabilities, but diagnostic reasoning, patient communication, and professional judgement. Medical education must evolve accordingly. This transformation resembles a “house moving“ rather than simple curricular addition: existing competencies must be rigorously evaluated, discarding those obsolete by algorithmic advancement; preserving those whose fundamental importance transcends technological change; and deliberately integrating new competencies essential for AI-enhanced healthcare delivery. E.g., which diagnostic skills retain irreplaceable value when algorithms can analyse imaging with superhuman accuracy? Which forms of clinical reasoning must we preserve as AI transforms information access? These questions demand ongoing attention. This process must remain organic and adaptive, as competency requirements will evolve continuously rather than stabilize. The integration of AI into medical education is inextricably linked to its implementation in clinical practice. Healthcare AI adoption will likely follow a non-linear path, constrained by data availability, workforce readiness, and infrastructure limitations that create friction between rapid technological development and measured clinical deployment. Medical curricula must therefore prepare students for an AI-integrated future while acknowledging the uneven, incremental reality of healthcare transformation. This competence review should be tackled from five major fronts:Preclinical competencies: subjects such as biology, anatomy and physiology form the backbone of all later clinical development, relying heavily on memorization. Will AI substitute that effort? Which knowledge remains non-negotiable, to be held in the physician’s own mind rather than retrieved on demand? And how can these tools instead reinforce long-term retention and the transfer of foundational knowledge into practice?Clinical competencies: Technical education should serve clinical judgement, not supplant it. Bedside skills, communication, and reasoning under uncertainty remain the human core of medicine. Certain examination abilities deserve protection precisely because technology can fail, or because patients need presence rather than mere conclusions.Research and health data science competencies: Future physicians must critically appraise studies designed, analysed, and even written with AI. Evaluating evidence now includes scrutinizing the algorithms behind it. Also, they should be provided with strong basic notions about health data science23, as it is AI's main foundation.Ethics competencies: No physician would prescribe vitamin A derivatives without first asking whether the patient is pregnant. AI use demands identical discipline. Technical fluency alone is insufficient; the deontological code that applies to other medical tools or treatments must prevent the ethical impacts of algorithmic tools.Technical competencies: they represent the genuinely novel frontier, absent from traditional curricula but increasingly essential. Medical students require sufficient technical literacy to practice intelligently with AI. Still, medical school cannot become computer science training. The curriculum is already overburdened; adding new content demands ruthless prioritization. The goal is not to produce engineers but to cultivate informed physicians who can use these tools wisely without needing to build them. Capacitation in this domain should include both faces of the coin: general technical AI knowledge and piloting of medical devices with AI. Preclinical competencies: subjects such as biology, anatomy and physiology form the backbone of all later clinical development, relying heavily on memorization. Will AI substitute that effort? Which knowledge remains non-negotiable, to be held in the physician’s own mind rather than retrieved on demand? And how can these tools instead reinforce long-term retention and the transfer of foundational knowledge into practice? Preclinical competencies: subjects such as biology, anatomy and physiology form the backbone of all later clinical development, relying heavily on memorization. Will AI substitute that effort? Which knowledge remains non-negotiable, to be held in the physician’s own mind rather than retrieved on demand? And how can these tools instead reinforce long-term retention and the transfer of foundational knowledge into practice? Clinical competencies: Technical education should serve clinical judgement, not supplant it. Bedside skills, communication, and reasoning under uncertainty remain the human core of medicine. Certain examination abilities deserve protection precisely because technology can fail, or because patients need presence rather than mere conclusions. Clinical competencies: Technical education should serve clinical judgement, not supplant it. Bedside skills, communication, and reasoning under uncertainty remain the human core of medicine. Certain examination abilities deserve protection precisely because technology can fail, or because patients need presence rather than mere conclusions. Research and health data science competencies: Future physicians must critically appraise studies designed, analysed, and even written with AI. Evaluating evidence now includes scrutinizing the algorithms behind it. Also, they should be provided with strong basic notions about health data science23, as it is AI's main foundation. Research and health data science competencies: Future physicians must critically appraise studies designed, analysed, and even written with AI. Evaluating evidence now includes scrutinizing the algorithms behind it. Also, they should be provided with strong basic notions about health data science23, as it is AI's main foundation. Ethics competencies: No physician would prescribe vitamin A derivatives without first asking whether the patient is pregnant. AI use demands identical discipline. Technical fluency alone is insufficient; the deontological code that applies to other medical tools or treatments must prevent the ethical impacts of algorithmic tools. Ethics competencies: No physician would prescribe vitamin A derivatives without first asking whether the patient is pregnant. AI use demands identical discipline. Technical fluency alone is insufficient; the deontological code that applies to other medical tools or treatments must prevent the ethical impacts of algorithmic tools. Technical competencies: they represent the genuinely novel frontier, absent from traditional curricula but increasingly essential. Medical students require sufficient technical literacy to practice intelligently with AI. Still, medical school cannot become computer science training. The curriculum is already overburdened; adding new content demands ruthless prioritization. The goal is not to produce engineers but to cultivate informed physicians who can use these tools wisely without needing to build them. Capacitation in this domain should include both faces of the coin: general technical AI knowledge and piloting of medical devices with AI. Technical competencies: they represent the genuinely novel frontier, absent from traditional curricula but increasingly essential. Medical students require sufficient technical literacy to practice intelligently with AI. Still, medical school cannot become computer science training. The curriculum is already overburdened; adding new content demands ruthless prioritization. The goal is not to produce engineers but to cultivate informed physicians who can use these tools wisely without needing to build them. Capacitation in this domain should include both faces of the coin: general technical AI knowledge and piloting of medical devices with AI. None of this materializes without prepared teachers. Faculty cannot transmit what they have not mastered. Their development is not a parallel effort but a precondition, the front upon which all others depend. These five fronts can be operationalized into learning outcomes, developmental milestones and assessment indicators; Table 1 provides illustrative examples, which should be adapted to local context. Nonetheless, enlisting such competencies demands caution. They should part from established foundational cores already embedded in medical curricula, such as those of the Association of American Medical College (AAMC)24 and the Accreditation Council for Graduate Medical Education (ACGME)25 and integrated with the AI-specific frameworks now emerging. Some examples for inspiration include the Delphi-derived competency sets of Lee et al. for South Korean graduates26 and Singla et al. for a Canadian curriculum27.Table 1Illustrative learning outcomes, developmental milestones and assessment indicators for AI integration across five curricular domains in undergraduate medical educationDomainIllustrative Learning outcomeDevelopmental milestoneAssessment indicatorPreclinicalRetain foundational knowledge required for reasoning without retrievalDistinguishes knowledge that must be internalized from what may be looked upOral or another isolated method of examination of core conceptsClinicalPerform and interpret core clinical tasks independently of AIExecutes key physical anamnesis, physical examination and key basic procedures (e.g. Cardiopulmonary resuscitation) and reasoningDirectly observed examination without AI accessResearch & Health Data ScienceCritically evaluate AI-assisted studies and grasp basic data-science principlesIdentifies methodological algorithmic limitations in a published AI studyOral or another isolated method of examination, so that unaided reasoning can be provedCritically conduct AI-assisted studiesDesigns, runs and reports a study using AI tools while identifying their methodological, ethical and algorithmic limitationsStructured evaluation of a student-led AI-assisted study, judged on transparency of AI useEthicsApply professional and deontological standards to AI useRecognizes bias, accountability, and consent issues in a clinical AI scenarioCase-based ethical analysis with justificationTechnicalUse and pilot medical AI tools safely without building themSelects, prompts, and validates an AI tool appropriately for a taskSimulated AI-assisted task with predefined tools and scoring criteria, judged on safe and effective use by evaluators who have mastered both the task and its AI-assisted performance. Table 1Illustrative learning outcomes, developmental milestones and assessment indicators for AI integration across five curricular domains in undergraduate medical educationDomainIllustrative Learning outcomeDevelopmental milestoneAssessment indicatorPreclinicalRetain foundational knowledge required for reasoning without retrievalDistinguishes knowledge that must be internalized from what may be looked upOral or another isolated method of examination of core conceptsClinicalPerform and interpret core clinical tasks independently of AIExecutes key physical anamnesis, physical examination and key basic procedures (e.g. Cardiopulmonary resuscitation) and reasoningDirectly observed examination without AI accessResearch & Health Data ScienceCritically evaluate AI-assisted studies and grasp basic data-science principlesIdentifies methodological algorithmic limitations in a published AI studyOral or another isolated method of examination, so that unaided reasoning can be provedCritically conduct AI-assisted studiesDesigns, runs and reports a study using AI tools while identifying their methodological, ethical and algorithmic limitationsStructured evaluation of a student-led AI-assisted study, judged on transparency of AI useEthicsApply professional and deontological standards to AI useRecognizes bias, accountability, and consent issues in a clinical AI scenarioCase-based ethical analysis with justificationTechnicalUse and pilot medical AI tools safely without building themSelects, prompts, and validates an AI tool appropriately for a taskSimulated AI-assisted task with predefined tools and scoring criteria, judged on safe and effective use by evaluators who have mastered both the task and its AI-assisted performance. Illustrative learning outcomes, developmental milestones and assessment indicators for AI integration across five curricular domains in undergraduate medical education Rethinking evaluation methods Parallel to competency review, evaluation methodologies must be fundamentally reconsidered. Each competency requires individual scrutiny. There will be no philosopher's stone, no unique answer. Yet one essential question must guide the process: Must future physicians perform this skill independently of AI assistance? This simple question carries profound implications that demand honest confrontation. Consider a straightforward example: cardiopulmonary resuscitation (CPR). Few would contest that physicians must perform CPR without algorithmic support. Nonetheless, this clarity dissolves when applied across the broader spectrum of clinical skills. The CPR example nevertheless illuminates a critical principle: if competency requires AI-independent performance, evaluation cannot occur in AI-accessible contexts. This constraint introduces significant pedagogical challenges. AI-free evaluation environments will become increasingly scarce as these technologies become embedded not only in conventional digital devices but also in wearable and environmental technologies, including smart glasses. Consequently, traditional assessment methodologies risk obsolescence. Authentic evaluation will also require clear expectations, professional integrity, direct observation, and assessment designs that make students’ reasoning processes visible. For competencies where AI assistance is appropriate, assessment criteria must be reconceived. The evaluative question transforms: not “Can the student solve this problem?” but “Can the student effectively collaborate with AI while maintaining clinical judgement?” These demands assessing prompt engineering, output validation, algorithmic limitation recognition, and the synthesis of human expertise with computational support. At the individual level Professional responsibility for AI literacy AI literacy is not only an educational opportunity but also an emerging professional duty for students and educators. Safe use requires sufficient understanding of a tool’s indications, limitations, potential harms, and appropriate conditions of use13. Students should progress from passive users to accountable professionals who critically appraise AI outputs, recognize inappropriate use, and retain responsibility for decisions informed by them. Educators should model transparent practice, acknowledge uncertainty, and maintain human oversight18. Engagement and transparency in AI-supported learning The educational value of generative AI depends not only on output quality but also on how openly and actively students engage with it. Students should understand its purpose, disclose its contribution when appropriate, and remain intellectually engaged in the learning process18. Transparent use allows educators to distinguish legitimate assistance from substitution of effort and enables meaningful feedback on how students formulate questions, evaluate outputs, and integrate information. AI should therefore support active learning rather than hindering how knowledge and conclusions were produced28. Empowerment, initiative, and creativity Students' capacity to learn, reason, and create independently should be protected. Students must be empowered to decide when AI adds value or weakens learning, rather than following the path of least cognitive resistance29. The objective is to master AI without becoming mastered by it: using the technology to extend curiosity, initiative, and creative exploration while retaining the confidence and competence to work without its assistance. Educational practice should engage original questioning, independent problem-solving, and thoughtful experimentation. This requires particular attention to the risks of cognitive offloading, dependency, and algorithmic overconfidence, as well as to AI’s still-uncertain effects on long-term memory and foundational clinical reasoning30,31. Policy and educational implications Policies supported by this framework include institutionally licensed AI environments, controlled pilots, and separate AI-free and AI-supported assessments. These approaches can improve equity, data protection, controlled innovation and preservation of independent competence, but require funding, infrastructure, monitoring and assessment redesign. Institutions should adapt them to local capacity without weakening non-negotiable safeguards. Infrastructure and cybersecurity AI implementation can expose educational and health information systems to privacy, security, and integrity risks affecting students, patients, research participants, and institutional partners14. Institutions should use approved environments with access controls, data minimization, logging, and security review, and keep sensitive data out of unapproved external models. Secure computational sandboxes can isolate experimentation from operational systems and confidential data. Interdisciplinary collaboration Responsible integration requires collaboration among medical educators, clinicians, computer scientists, cybersecurity and data-protection specialists, ethicists, and learners3,32. These partnerships should guide tool selection, evaluation, and co-design to ensure technical robustness, pedagogical value, and clinical relevance. Challenges to implementation Comprehensive governance requires more than institutional enthusiasm. Weak implementation structures, unstable funding, and unequal access may widen disparities between institutions16. Proprietary systems can restrict transparency and independent scrutiny3,14. Governance must therefore remain iterative, with policies, risk assessments, and oversight mechanisms reviewed as technologies and evidence evolve. The framework’s feasibility is bounded by resource constraints, and these vary sharply between institutions. Evidence from low-resource settings is clear: binding barriers are rarely algorithmic. They are infrastructural8. Unreliable electricity, intermittent connectivity, and ageing hardware routinely defeat tools that perform well elsewhere. Data pose a second barrier. Where records remain paper-based or fragmented, the structured digital substrate that fuels AI simply does not exist. Namely, a new perfect AI programme that can predict Intensive Care Unit health outcomes cannot be reasonably implemented in a setting where vital signs are registered on paper. Workforce and financing complete the picture: overburdened staff that cannot validate these tools, and reliance on proprietary systems and short-term funding leaves pilots to lapse once external support ends. Crucially, these constraints fall unevenly, widening the gap between well and under-resourced colleges8,33. Governance must recognize and adapt to this reality. Institutions must engage stakeholders in the process and perform a sincere mapping of needs. Tackling these barriers requires a two-front strategy: secure or better the foundations if possible (e.g. Power, connectivity, digital records…); and creative solutions that engage the community (e.g. Scheduling compute-heavy tasks, as model runs and downloads; during stable supply windows, Offline AI tools with local data catching; preferring open-source tool and sustainable and long-term public-private partnerships; a “train the trainer” philosophy that cascades AI literacy, that invites both educators and students to learn and validate tools; consortia with local technical colleges that could benefit and supply technical skills and maintenance; and progressive migration to interoperable digitalization of data, in non-proprietary information systems). Conclusions AI governance in undergraduate medical education must urgently move beyond reactive compliance towards proactive capacity building. This requires coordinated action across institutional, curricular, and individual levels, with clear accountability, shared responsibility, and safeguards proportionate to educational risk. Governance should preserve competencies that require independent performance while enabling transparent and critically supervised AI-supported learning and academic monitoring. Unequal access, lack of protected training time, funding and the need for a multidisciplinary approach are challenges that must be strategically addressed. As a conceptual and untested framework, its feasibility, acceptability and effects on learning, equity and safety require empirical evaluation across institutions with differing resources and digital maturity. Its ultimate purpose is to ensure that AI strengthens clinical judgement, learner agency, and patient-centred care.