AI is rapidly transforming medical education beyond the use of conversational agents such as ChatGPT. Its current impact can be categorized into four main areas: interactive learning tools, intelligent assessment systems, administrative and logistical support, and content creation.

AI-driven virtual assistants and chatbots—such as ChatGPT—facilitate dialogue-based learning, allowing students to practice history taking and enhance communication skills through simulated patient encounters [6]. Moreover, immersive technologies, including virtual reality, intelligent tutoring systems (ITS), medical robotics, and augmented reality, provide hands-on training experiences. These tools promote the development of clinical reasoning, surgical technique, and medical imaging interpretation while delivering personalized feedback and adaptive learning pathways [7,8].

AI enhances evaluation processes by enabling automated grading and reducing evaluator bias. NLP can analyze narrative feedback to identify patterns related to competency, professionalism, and potential learner risk [9]. Innovations such as virtual Objective Structured Clinical Examinations (OSCEs) and automated assessments of clinical case presentations streamline performance evaluations [10-13]. In addition, ML contributes to the creation and validation of examination items, although human oversight remains essential to maintain content integrity [14,15].

In educational administration settings, AI automates the documentation of clinical experiences. For example, NLP applications can map trainee documentation to core competencies with high accuracy (92%‐97%), significantly reducing clerical workload [16-19]. Furthermore, predictive modeling aids in optimizing residency selection and procedural tracking, enhancing the efficiency of academic programs [20,21].

AI also supports academic writing by assisting in the drafting of structured clinical notes, research manuscripts, and literature summaries. These tools enhance clarity, reduce writing time, and facilitate effective scholarly communication [22,23] (Table 1).

Despite its benefits, the incorporation of AI into medical training poses several concerns. One is the potential impact on career decisions. Some students may be deterred from entering fields such as radiology due to perceived reductions in future demand, as AI systems become increasingly proficient at image interpretation [29]. Another concern is the limited integration of AI-related content in existing medical curricula, leaving students unprepared to engage with AI-driven health care systems [30]. Additionally, there is the risk of overreliance on AI tools, which may undermine the development of critical hands-on clinical skills, diagnostic reasoning, and human-centered decision-making [31].

The ethical use of AI in medical education is a growing focus of scholarly and institutional discourse. A key consideration is the preservation of human judgment. While AI can support decision-making, it must not displace the clinician’s responsibility for critical thinking, empathy, and ethical deliberation [32]. Moreover, transparency and accountability are essential. Medical students must be taught to recognize biases, uphold data privacy, and ensure fairness in AI-mediated outcomes. Equitable access is also critical; AI integration must address the digital divide to ensure that learners across different regions and socioeconomic backgrounds benefit equally [29,33,34].

In medical education, AI use may compromise several ethical principles: academic integrity (fabrication or inaccuracy of references), justice (potential biases), nonmaleficence (overreliance on algorithmic outputs by students), and confidentiality (risks of training data extraction). These concerns are supported by empirical findings. For example, in a set of 115 references generated with ChatGPT-3.5 (OpenAI), 47% (n/N) were fabricated, 46% (n/N) inaccurate, and 7% (n/N) authentic and correct [35]; in another analysis of 636 citations, 55% (n/N) were fabricated with GPT-3.5 and 18% with GPT-4 (OpenAI), alongside errors in “real” citations [36]. Editorial practice has likewise documented high similarity indices with plagiarism risk and failures to detect misuse in medical-scientific articles—including AI-generated images with biological errors and the publication of stock large language model (LLM) phrases (eg, “I do not have access to real-time information...”) in papers that passed peer review [37]. Beyond GenAI, computer vision used for instructional purposes (eg, medical imaging or intraoperative video) exhibits performance degradation due to dataset shift, dependence on spurious signals across hospitals or equipment, and vulnerability to adversarial attacks, necessitating external validation, audits, and technical safeguards when incorporated into educational activities [38,39]. From a privacy standpoint, LLMs have demonstrated memorization of training data and the potential for verbatim disclosure, reinforcing the need to avoid uploading protected health information (PHI) and to adopt institutional controls (data policies, privacy-preserving environments, and prior review before use in classrooms or OSCEs) [40,41].

In alignment with the United Nations Educational, Scientific and Cultural Organization (UNESCO) Recommendation on the Ethics of Artificial Intelligence (2021) and the World Health Organization Guidance on Ethics and Governance of AI for Health (2021), we translate principles into practice through five tightly scoped, actionable axes [42,43]: (1) governance and accountability: establish an oversight committee, maintain a public registry of AI systems, and require predeployment Algorithmic Impact Assessments and Data Protection Impact Assessments; (2) human oversight in high-stakes uses: mandate dual human+AI scoring, structured debriefs, and explicit override criteria grounded in clinical or educational judgment; (3) transparency: disclose in syllabi which tools are used and for what purposes and limits, provide model or data cards, and require student disclosure of AI assistance; (4) privacy-preserving data governance: prohibit uploading PHI to public models; enforce data minimization and deidentification, role-based access controls, secure logging, and retention limits; and prefer on-prem and virtual private cloud solutions when processing student data; and (5) equity, safety, and robustness: conduct subgroup bias testing and disparate impact monitoring, pursue cross-site validation where feasible, and set predefined error thresholds with rollback plans. To operationalize these pillars, we propose a staged pathway that begins with pilots in controlled sandbox environments with Algorithmic Impact Assessments or Data Protection Impact Assessments and a clear evaluation plan, then scales gradually with bias and robustness audits and cross-site performance monitoring, and culminates in institutional integration with periodic audits, update protocols, and public reporting. Together, these steps aim to improve learning while safeguarding autonomy, justice, and trust.

To ensure responsible and effective use of AI, medical education should incorporate foundational and applied competencies. Core knowledge areas should include data science, ML, algorithm design, statistics, and basic coding principles [44,45].

Curricular modules should also address clinical applications of AI in domains such as surgical training, radiology interpretation, ophthalmologic analysis, and hematologic diagnostics [46-49]. Ethical and legal education remains vital, with an emphasis on transparency, privacy, and professional accountability [34,50]. Notably, one publication has even proposed establishing a dedicated specialty in Medical Data Sciences to address the growing demand for AI expertise in clinical settings [51].

Educators are actively exploring pedagogical strategies to integrate AI into undergraduate medical training. AI-enhanced curriculum design, such as the use of chatbots and generative platforms, enables the creation of dynamic learning experiences tailored to student progression and clinical reasoning levels [28]. Similarly, adaptive assessment systems powered by AI can identify individual student weaknesses and provide targeted feedback, thereby optimizing skill development and educational outcomes [23,27].

Despite increasing interest and promising use cases, several challenges continue to hinder the effective integration of AI in medical education. Ethical and legal concerns include risks to data privacy, opaque accountability for AI-supported decisions, and the potential erosion of humanistic principles in teaching and care; AI use must remain aligned with the ethical foundations of medical practice and must respect the clinician-patient relationship [53]. Unreliable or low-quality outputs from generative models, such as ChatGPT and similar tools, may perpetuate outdated, biased, or inaccurate information when these systems are trained on flawed data; inaccuracies in AI-generated educational content can compromise learner safety and clinical competence, underscoring the need for rigorous data curation and active human oversight [1,54]. Scalability and equity barriers arise from limited digital infrastructure and high implementation costs, particularly in low- and middle-income countries, where disparities restrict equitable access to AI-driven tools and may reinforce existing educational inequities [54]. Finally, curricular displacement is a concern: overdependence on AI may inadvertently marginalize foundational clinical reasoning, communication, and hands-on skills if these technologies are introduced without explicit pedagogical safeguards and thoughtful alignment with existing curricula [51].

To harness the benefits while mitigating potential risks, a comprehensive strategy is essential—one that encompasses ethical considerations, pedagogical advancements, and technological inclusivity. Several complementary approaches can facilitate the responsible adoption of AI in medical education. Strengthening ethical frameworks is critical to ensure transparency, fairness, and respect for patient-centered values throughout AI development and deployment [53]. Enhancing data quality and human oversight is equally important: training AI systems on peer-reviewed data and subjecting outputs to expert review can reduce misinformation and support safe implementation [1,54]. Promoting equitable access involves designing adaptable, low-cost solutions tailored to diverse educational contexts and resource environments, particularly in underserved regions [54]. Reforming curricula to include AI literacy prepares future physicians to engage critically with algorithmic tools, with core competencies that encompass data science, algorithmic reasoning, ethical discernment, and digital fluency [51]. Supporting longitudinal research on the long-term educational and clinical impacts of AI integration will provide the evidence base needed for informed decision-making and continuous improvement [54].

These strategies are further detailed in Table 2, which aligns specific challenges with corresponding solutions to guide effective AI integration in medical education.

In the current context, recent studies typically address isolated components of AI in medical education, such as specific teaching technologies or environments. However, to advance this field, a more cohesive framework is required that facilitates replication, innovation, and a comprehensive understanding of the educational role of AI [2]. To this end, the FACETS framework has been proposed as a guiding model for future research and implementation in medical education settings. By examining six key dimensions—form, application, context, instructional mode, technology, and the SAMR (substitution, augmentation, modification, redefinition) model—this framework provides a structured approach for assessing how AI tools align with educational objectives (Figure 1). This multifaceted analysis ensures that AI implementations are pedagogically sound and contextually appropriate [2].

The tangible applicability of FACETS can be appreciated by retrospectively mapping published interventions (Table 3). Holderried et al [55] evaluated a GPT-4–based simulated patient that provided immediate, structured feedback for history taking among third-year medical students; they observed more than 99% medically plausible exchanges and near-perfect agreement with a human rater (κ=0.832), although 8 of 45 feedback categories showed lower concordance, underscoring the need for human oversight. Similarly, Luordo et al [56] used GPT-4 to grade OSCE reports from 96 students, finding high concordance with human graders (intraclass correlation coefficient=0.77 for single measures; 0.91 for average measures), systematically stricter AI scoring (AI mean 8.88, SD 2.96 vs experts’ mean 12.39, SD 3.22; mean difference −3.51 points), and substantially shorter grading times (~24 min vs 2‐4 h). Taken together, although the original studies did not explicitly use FACETS, this mapping supports the utility of the FACETS framework as an analytic lens to describe, compare, and guide future implementations of AI in medical education, particularly in undergraduate contexts.

Beyond the specific cases mapped earlier, FACETS can also be used to anticipate and assess emerging applications of AI in medical education—ranging from ITS and personalized learning platforms to chatbots (eg, ChatGPT) and intraoperative video analysis—by aligning each use case with pedagogical intent, context, and level of transformation [2].

As AI becomes more deeply embedded in health care delivery, its incorporation into medical education continues to represent a critical yet uneven frontier. Although AI tools have shown promise in diagnosis, therapeutic planning, and operational efficiency, educational adoption has lagged—both in scale and in the rigor of implementation and evaluation. Key barriers include heterogeneous technological infrastructure—particularly in resource-limited settings—and limited AI literacy among students and faculty [57,58]. These challenges are compounded by the fact that many medical students, despite their interest in AI’s potential, express anxiety about downstream labor market implications [59,60].

Beyond the case-based exemplars mapped with FACETS, diverse emerging applications in medical education share several salient features: they prioritize deliberate practice with adaptive feedback and competency-aligned performance traceability; they operate predominantly at the augmentation and modification levels of the SAMR model (with few achieving true redefinition); and they rely on high-quality data (interaction logs, rubrics, audio-video) that demand psychometric validation, bias and drift monitoring, and robust privacy safeguards. The evidentiary base remains heterogeneous, with single-site pilots and intermediate metrics predominating over longitudinal outcomes. The human factor remains essential—explicit instructional design, structured debriefing, and human-in-the-loop approaches are needed to calibrate criteria and prevent overreliance. Absent targeted faculty development and institutional processes (including data governance, audits, and reporting standards), these solutions risk amplifying equity gaps.

To operationalize these principles and translate them into programmatic, monitorable curricular decisions, it is helpful to note that these common features are especially evident in intelligent tutoring systems (ITSs), clinical chatbots, AI-assisted assessment, personalized learning platforms, virtual reality and AI simulators, anatomy tools, and intraoperative video analytics (Table 4). Building on that foundation, cross-cutting applications that enable and govern implementation across workstreams assume particular importance: (1) multimodal adaptive analytics, which fuse text, audio-video, and interaction traces to personalize pacing and content while flagging early risk at the learner or course level [61]; (2) learning analytics dashboards, which integrate competencies, assessment evidence, and progression signals to guide just-in-time feedback, remediation, and coaching for students and instructors [62]; and (3) educational clinical decision support, implemented in deidentified, guideline-constrained sandbox environments to teach evidence retrieval, uncertainty appraisal, and human-AI teaming in diagnostic and therapeutic planning [63,64] (Table 4).

Effective adoption of these cross-cutting layers requires the very institutional capacities and safeguards they presuppose. Curricula must evolve to incorporate AI training that fosters both technical competence and ethical responsibility [2]. This entails not only understanding AI tools but also recognizing their limitations, biases, and societal implications. Cross-sector collaboration among educational institutions, technology developers, and researchers is crucial to establishing universal digital literacy standards that equip learners and faculty to critically evaluate AI-generated content and apply such tools responsibly in clinical and educational settings [4].

AI-powered platforms are expanding personalized learning by tailoring educational content to individual needs. Technologies such as virtual patients, augmented reality simulations, and mobile platforms offer dynamic, interactive experiences that increase engagement and broaden access to high-quality education, particularly in resource-limited regions. Nonetheless, the risk of depersonalizing medical practice—together with substantial concerns about data privacy—underscores the need for rigorous ethical and regulatory frameworks that engage all stakeholders in the educational process [75].

Given AI’s expanding role in health care, prohibiting its use in education is neither practical nor beneficial. Institutions should instead establish comprehensive guidelines to ensure that AI-generated content remains reliable, relevant, and ethically sound. A structured, responsible approach—grounded in robust educational infrastructure, interdisciplinary collaboration, and consistent regulatory oversight—has the potential to improve learning outcomes and, ultimately, strengthen patient care.

As a synthesis of the manuscript’s arguments, we recommend a concise, actionable set of practices: (1) report every AI pilot with a FACETS-aligned template to ensure transparency and comparability; (2) maintain human oversight for high-impact uses (eg, dual scoring with override), reserving automation-first approaches for low-risk contexts with active monitoring; (3) conduct predeployment audits for bias, drift, and adversarial robustness and, where appropriate, cross-site validation before scaling; (4) operate privacy-preserving workflows (no PHI in public LLMs, institutional sandboxes, data minimization or deidentification, role-based access, secure logging and retention); (5) deploy learning analytics dashboards tied to competency frameworks with ongoing psychometric monitoring; and (6) evaluate equity and feasibility using implementation science designs, including in resource-limited settings with disparate impact monitoring. This roadmap is summarized in Figure 2.