Based on the theoretical framework of symbiotic agency [11], a theory emphasizing interdependent and collaborative relationships between humans and technology, this study conceptualizes human-technology relations as a process of mutual constitution. Technology functions neither as a passive instrument dominated by humans nor as an autonomous replacement for human agency. Instead, it develops in tandem with humans through interdependent interactions: technology enhances human efficacy by expanding cognitive boundaries and enabling novel multimodal interactions, while humans legitimize technological practice by embedding ethical norms and conducting context-specific interpretations such as weighting clinical decisions. This symbiosis transcends traditional master-servant dichotomies by establishing a responsibility-sharing network. Within this network, technology acts as a co-agent in human activity systems, collectively enhancing capabilities rather than substituting human roles. This perspective provides the foundational understanding needed to maintain a dynamic balance in human-technology interdependence within medical education, forming the basis of our conceptual model.

Building upon the analytical framework established in Table S1 in Multimedia Appendix 1, which systematically compares general-purpose and domain-specialized GAI models across 3 critical dimensions (knowledge representation fidelity, task compatibility, and ethical constraint mechanisms), this study deconstructs technological heterogeneity to avoid conflating “GAI” as a homogeneous entity. The models in Multimedia Appendix 1 (see Table S1) were selected via multisource evidence synthesis, including peer-reviewed studies, industry reports (eg, Global Large Language Model (LLM) Market Research Report 2024) [12-20], and empirical validation in educational contexts, based on four criteria: (1) technological representativeness of core advancements (multimodality, reasoning, and domain adaptation); (2) broad academic and practical relevance in medical education; (3) functional diversity covering text, image, video, and domain-specific tasks; and (4) market prevalence, wide recognition, technical maturity, and development by prominent AI companies. Notably, models like Perplexity, DeepSeek, Notebook LM, and Midjourney, though used by clinicians and students in specific scenarios, were not included due to limited evaluative data and insufficient supporting information in the referenced reports.

Within this ecosystem, general LLMs serve as multitasking hubs, leveraging cross-domain adaptability and natural language interaction, while specialized models achieve context-specific efficacy through the embedding of deep medical knowledge. To resolve their complementary yet fragmented coexistence, we propose a specialized model integration system anchored to general LLMs, inspired by symbiotic agency theory and hospital diagnostic workflows [21] (see Figure 1). This architecture establishes a 3-tiered clinical analog: general LLMs serve as primary coordinators, managing task orchestration; specialized models act as domain experts, executing depth-specific processing; and protocol-based collaboration enables online consultation through knowledge distillation and output validation. This hierarchical integration embodies symbiotic agency principles: general models extend the applicability of specialized techniques by transcending domain boundaries, while specialized models enhance system depth by reinforcing medical logical rigor. Through functional complementarity and role differentiation, they form a synergistic symbiont exceeding individual capability limits, establishing an intelligent foundation for medical education characterized by adaptability, expertise, and reliability.

The tripartite synergy paradigm, rooted in complex systems management theory and evidenced across domains from political governance to integrated health care systems (eg, the mission alignment model by Peek et al [22]), establishes our resource-method-assessment (RMA) framework (see Figure 2) as the core analytical structure [22,23]. This framework defines three interdependent dimensions: (1) resources encompassing dynamic content provisioning mechanisms, (2) methods designing knowledge-to-practice training pathways, and (3) assessment managing outcome monitoring and feedback generation. Their cyclical optimization forms an integrated whole, as resource renewal enables pedagogical innovation, method implementation yields evaluative data, and assessment outputs drive resource refinement and method calibration. Within this architecture, GAI operates as a collaborative instrument executing content generation, interaction support, and data analysis under educator-directed goal design, ethical governance, and critical intervention. The established framework provides essential categorization criteria for subsequent empirical analysis: it defines 3 dimensions—resources, methods, and assessment—directly corresponding to 3 primary research domains in GAI applications for medical education. By consolidating fragmented literature within a unified analytical structure, this framework systematically addresses cognitive limitations arising from isolated examinations of technological functions, thereby elucidating the intrinsic operational logic of technology-enabled educational transformation.

Before comprehensive data extraction, a structured data extraction form was collaboratively developed by all authors. This iterative process was guided by our research questions and the predefined thematic framework outlined in Table 1, which focused on the applications, challenges, and prospects of GAI in medical education applications. The form was designed to systematically capture key information from each included article, encompassing: bibliographic details (eg, authors, publication year, journal, and country or region), study characteristics (eg, research design, objectives, and population), specific GAI models used (eg, ChatGPT [OpenAI] and Gemini [Google]), application scope (single-model vs multimodel), analysis type (performance comparison across models or examination of synergistic enhancement through model integration), detailed descriptions of identified applications, challenges, and future directions of GAI application in medical education categorized exclusively through our tripartite Trinity Framework and quantitative performance metrics (reported accuracy rates, percentages, mean scores, standard deviations, and P values related to GAI model performance in various tasks). This granular definition of data points ensured that all relevant information pertinent to our broad research inquiry was systematically collected.

To better understand the global research landscape in this field, we analyzed the countries or regions of origin for the 131 selected articles. For those without a precise location, we assigned them according to the country or region of the corresponding author’s institution. To analyze the distribution of research based on the countries or region’s development level, we used the Human Development Index (HDI) classification. The latest HDI data categorizes countries or regions into 4 tiers: very high, high, medium, and low human development with higher HDI scores correlating with greater national development. We also investigated cross-level HDI collaborations, which refer to partnerships between countries from different HDI categories [25].

To validate the comprehensiveness and clarity of the data extraction form, a pilot test was independently conducted by 2 reviewers, YL and ZL, on a randomly selected subset of 10 included articles. During this pilot phase, any discrepancies in data extraction or ambiguities within the form were identified and discussed. Based on these discussions, the data extraction form underwent iterative revisions to refine categories, clarify definitions, and ensure consistent interpretation of data points among reviewers. Following this refinement, YL and ZL independently extracted data from all 131 included articles. In cases of disagreement between the 2 independent extractors, consensus was initially sought through discussion. If a consensus could not be reached, a third and fourth reviewer, ZY and NZ, were consulted to mediate and make final determinations regarding the applicability and extraction of the data.

The performance of different question types: The studies encompassed a range of question types, including multiple-choice questions (MCQs), single-choice questions, short-answer questions (SAQs), true or false questions, open-ended short-answer questions (SOAQs), long-answer questions, clinical case analysis questions (CAQs), and image-text integrated questions [29-37]. An exploratory study conducted by a research team from Qatar University evaluated ChatGPT’s performance across various assessment formats relevant to undergraduate dental education. The study included 50 assessment items covering 50 different learning outcomes, with 10 items for each of the 5 formats: MCQs, SAQs, short essay questions (SEQs), single true or false questions, and fill-in-the-blank items. These items were based on core clinical topics in dental education, such as restorative dentistry, periodontics, endodontics, and oral surgery, aligned with the learning outcomes expected of undergraduate dental students. In this study, ChatGPT demonstrated 90% accuracy for SAQs, SEQs, and fill-in-the-blank items and notably achieved 100% accuracy in single true or false questions [31]. However, other studies have revealed a significant decline in accuracy for CAQs, as low as 17%, which require strong logical reasoning and lack predefined options [34].

Regarding MCQs, a study reported that GPT-4 and Microsoft Bing achieved top scores (76%) on the University of Antwerp medical licensing MCQ exam, outperforming medical students. However, ChatGPT’s accuracy fell considerably when tackling Chinese-language medical MCQs with an accuracy of 37%. In addition, another study reported that in the Chinese Master’s Degree Entrance Examination, ChatGPT’s accuracy for single-choice questions (A1 type) was 56%, whereas for MCQs, it dropped to 33% [15]. These findings suggest that GAI’s performance is not uniformly robust across all MCQ types and is influenced by factors such as question structure, subject domain, difficulty, language, and the presence of clinical vignettes or images.

The performance of different difficulty questions: In terms of difficulty, questions were generally classified as “easy,” “medium,” or “difficult.” For example, the difficulty levels are defined based on the performance indicators of the historical question bank: “Difficult” (P<.30; less than 30% of the students answered correctly), “Medium” (P=.30 to .80), and “Easy” (P>.80) [35]. ChatGPT-4 demonstrated strong performance on easy questions, with accuracy rates reaching 97.4%. Yet, even in this category, ChatGPT-4’s performance lagged behind that of residents [35,38-42]. In contrast, ChatGPT-4 excelled on medium and difficult questions, outperforming residents by 25.4 and 24.4 percentage points, respectively [38]. Across all models, performance tended to decline with increased difficulty, especially for higher-level questions that required multistep reasoning, where accuracy dropped markedly [39-45].

The performance of questions at different cognitive levels: 2 studies investigated the performance of ChatGPT-4 on questions categorized by Bloom’s taxonomy, which includes 6 cognitive levels: remembering, understanding, applying, analyzing, evaluating, and creating [46]. These studies found that ChatGPT-4 consistently performed well across all cognitive levels, with an average correct answer rate of 71.96% for each cognitive level [47,48].

By collaborating with instructors, GAI can quickly generate comprehensive clinical cases, including patient history, physical examination results, lab data, and differential diagnoses tailored to predefined learning objectives (eg, chest pain and joint pain). This reduces the time instructors spend developing such cases [49,50]. Furthermore, GAI-generated cases can integrate various contextual factors such as race, occupation, and lifestyle, significantly enriching the diversity of teaching materials [51]. For example, when creating a case based on a disease profile specific to a region, the ethnicity of the generated patient can be adjusted accordingly. In the context of type 2 diabetes, modifications can be made to the age range and weight distribution. In addition, randomized prompts for urine analysis may be included in urinary tract infection cases. Both patient presentations and examination findings can be randomized, and symptom expression can be customized to meet specific learning needs [51].

In Smith and colleagues’ [52] study, GAI was assigned the task of creating a case of an immigrant with mental health concerns, as this group may require specialized social psychiatry interventions. The results indicated that GAI was able to produce a case that met fundamental educational objectives. However, it included several signs of emotional disorders, highlighting a need for further refinement.

Studies have shown that GAI is effective in promoting interactive learning and providing practice in communication skills. GAI-powered simulation tools simulate changes in clinical conditions in scenarios such as advanced cardiac life support (ACLS) and intensive care unit (ICU) sepsis, prompting students to critically analyze whether their decisions are correct [53]. In addition, conversational GAI-created digital patients provide anesthetists with valuable training for patient interactions, reducing reliance on human actors while enhancing the flexibility and consistency of the training process [54]. These digital interactions create a safe space for repeated practice, providing dynamic learning experiences that traditional textbooks cannot match [52,55]. Furthermore, conversational GAI models, such as chatbots, can simulate the role of a professor, offering critical evaluations of literature and distilling complex research into easily understandable key findings, thus fostering simulated discussions between students and experts in the field [56]. However, besides experiences and qualitative observations, formal evaluations of the reliability and validity of such GAI-generated information in a professor-like capacity are still needed.

By generating accessible public health information, GAI enhances the public’s understanding of essential health issues, such as infectious disease prevention and vaccination, ultimately leading to improved health literacy [30]. Furthermore, GAI-generated clinical cases can be disseminated as Open Educational Resources (OERs), providing medical educators with globally adaptable teaching materials that are customized to local contexts [51].

GAI tools such as Adobe Firefly, DALL·E 2 (OpenAI), Bing Image Creator, and generative adversarial networks (GANs) can create clinical images displaying various pathological features based on textual descriptions, potentially addressing the shortage of authentic pathological images in traditional medical education due to medical confidentiality and patient privacy restrictions [51,57-59]. For instance, images of retinal disease generated by the stable diffusion model enhance students’ learning opportunities in ophthalmic pathology, greatly enhancing the availability of visual teaching resources [57]. However, their reliability and accuracy vary significantly across models and tasks. For example, DALL·E 2 demonstrated an overall clinical accuracy rate of 22.2% in aligning generated images with textual prompts across 15 semantic relations (eg, spatial and action-based relationships), with only 3 relationships (touching, helping, and kicking) achieving moderate consistency above 25%. In a medical education context, DALL·E 2 achieved 78% accuracy for soft-tissue tumor images but produced inconsistent results for wound images, with 65% of generated wound images containing anatomical inaccuracies or irrelevant elements [58]. A comparative study of DALL·E 2, Midjourney, and Blue Willow for generating skin ulcer images showed DALL·E 2 performed best with an average score of 3.2/5 (scale 1‐5) but still produced irrelevant content (eg, X-rays instead of pressure ulcers) in 20% of cases. Midjourney generated stylized, exaggerated features in 40% of images, while Blue Willow produced images with little relevance to prompts in 70% of attempts [59].

GAI shows potential in the early stages of curriculum development, aiding in quickly creating course objectives, learning strategies, and frameworks. For instance, in a study on integrated pharmacotherapy of infectious disease education modules, ChatGPT helped design curriculum goals (eg, “describe mechanisms of antibiotic resistance”) with an average expert rating of 92% for appropriateness and accuracy, supporting educators—especially in designing foundational courses [60].

Researchers have developed derivative applications based on classical models, such as Glass AI (a powerful AI-driven knowledge management system developed by Glass Health, focusing on organizing and retrieving health-related information efficiently). It integrates GPT-4 with evidence-based, peer-reviewed clinical guidelines to generate differential diagnoses and clinical plans based on textual input of clinical cases, enabling students to interact with it and experience the GAI-driven diagnostic process for cases [61]. Similarly, an MCQ generator based on ChatGPT-generated cases offers a dynamic platform for personalized learning assessments [62].

Research shows that when GAI is used to answer MCQs, the explanations generated by GAI can better convey key knowledge points and achieve good accuracy and degree of matching with teachers’ explanations. Of the 81 questions explained by the teacher and correctly answered by ChatGPT, 92.6% of the explanations were accurate and included at least part of the teacher’s explanation. However, the research also highlights that if an initial response is incorrect, the likelihood of subsequent errors increases significantly (P<.001), indicating that an early mistake may lead to systematic inaccuracies in later explanations [63]. Complementing this, in a systematic review, the broader literature reviewed showed that the majority of studies (5/8, 62.5%) indicate the effectiveness of AI in generating valid MCQs, with a preference for the latest GPT-4 models (6/8, 75%) [64].

Studies demonstrate that GAI boosts students’ learning efficiency across multiple stages by offering personalized feedback and customized content. This includes support for exam preparation [55,65-68], optimizing learning paths and review strategies [52,69-71], clarifying medical concepts [68,72-76], and assisting in the development of tailored career plans [77]. For instance, in physiological case analysis, GAI offers precise responses and contextually relevant feedback. A cross-sectional study tested 77 physiology case vignettes (covering diverse physiological and pathophysiological scenarios, designed for undergraduates) on ChatGPT 3.5, Google Bard, and Microsoft Bing. Rated by two physiologists on a 0‐4 scale, ChatGPT scored highest at 3.19 (SD 0.3), outperforming Bard (2.91, SD 0.5) and Bing (2.15, SD 0.6) with P<.001. ChatGPT’s precision accelerates task completion, helping students grasp medical knowledge in practical scenarios more effectively [78]. Furthermore, a study found that in cases of initial incorrect responses, GPT-4 was able to self-correct and provide accurate answers after simple follow-up questions or hints, mimicking pedagogical interactions observed in residency programs. This dynamic learning approach, coupled with rapid information processing, positions GPT-4 as an important asset for personalized learning [79].

GAI uses its ability to analyze complex, domain-specific knowledge to support the diagnosis of rare and intricate diseases. In addition to diagnosis, it can generate differential diagnoses tailored to the unique characteristics of each disease, providing health care professionals with precise decision-making support [80-83]. For common pathological issues and basic data analysis, GAI tools are efficient and accurate, helping pathologists organize their thought processes and expedite the initial diagnostic phases [84]. The impact of domain-specific training is profound. For instance, refined datasets in the surgical and anesthesiology fields enhance GAI’s clinical decision-making capabilities. In scenarios such as a “30-year-old pregnant woman requiring an emergency appendectomy,” GAI suggests not only tailored surgical strategies but also factors in critical anesthesia protocols [85]. Furthermore, in the field of traditional Chinese medicine, when combined with such tools, GAI can effectively create knowledge maps that organize entities, attributes, and their relationships to traditional Chinese medicine through graphical structures. GAI provides unique support for teaching traditional Chinese medicine and disease diagnosis and treatment decisions [86].

GAI demonstrates potential in multidisciplinary knowledge acquisition within medical education by providing high-quality knowledge across various medical subfields [87-94]. GAI demonstrates adaptability across disciplines, including shoulder and elbow surgery, sports medicine, and oncology [91]. Research further indicates that GAI models such as ChatGPT-4 excel in internal medicine, pediatrics, obstetrics and gynecology, surgery, emergency care, and public health [88-90,92-94]. Notably, a study assessing ChatGPT-4’s performance in the American Board of Family Medicine (ABFM) certification examination demonstrated its significant proficiency, with both the custom robot version (embedded in a specialized subenvironment designed to mimic examination conditions and given extensive preparation resources) and the regular version (standard ChatGPT-4) achieving high correct response rates of 88.67% and 87.33% respectively, well above the passing threshold. This further highlights GAI’s value in enhancing medical education within a multidisciplinary framework, making it a powerful learning support tool across a wide range of fields, including family medicine [95]. A meta-analysis of ChatGPT-3.5/4 across medical, pharmacy, dentistry, and nursing licensing exams revealed an overall accuracy of 70.1% (95% CI 65%‐74.8%; P<.001). Performance varied significantly by field (Q=15.334; P=.002), with pharmacy having the highest rate (71.5%, 95% CI 66.3%‐76.2%) and nursing having the lowest rate (61.8%, 95% CI 58.7%‐64.9%). These results demonstrate GAI’s potential to provide multidisciplinary learning support in health professions [96]. It is crucial to note that the evidence presented in this section highlights the individual learner’s ability to access and comprehend information across disciplines. This review’s existing evidence has not yet extensively covered GAI’s direct support for complex interdisciplinary teamwork, closed-loop communication, or the cultivation of specific professional behaviors within collaborative learning environments.

A study shows that GAI excels in creating article outlines and editing formatting, which alleviates common writing challenges related to poor organization and grammatical mistakes [28]. In addition, GAI can significantly improve the quality and standardization of academic writing, allowing medical educators and students to express their ideas more accurately and clearly [28,55,97]. Furthermore, GAI assists students in organizing and generating literature content while writing their thesis [98]. The content produced by GAI maintains consistency in language and includes appropriate academic terminology and logical structure, helping students present themselves more professionally in their academic writing [55]. Furthermore, GAI supports many non-native English speakers in overcoming language barriers during the academic writing process, which enables them to engage more confidently in academic communication [71].

Overreliance can be caused due to the following factors.

First is impaired critical thinking. The rapid feedback provided by GAI may reduce students’ time for deep thinking, weakening their ability to analyze problems and independently engage in critical learning. This phenomenon is particularly evident in medical education, where students often rely on the answers provided by GAI when solving complex problems rather than relying on their logical reasoning and knowledge accumulation for analysis and resolution [35,39,40,50,55,69,70,72,74,75,77,98,99,124,136,146-152].

Second is decreased creativity. When students use GAI tools like ChatGPT, they often receive writing suggestions that lack the creativity and depth of human-generated content. Thus, prolonged reliance on such tools may weaken their independent writing skills and hinder their ability to engage with complex topics that require critical thinking and practical expertise [28]. Similarly, educators who overly depend on GAI for content creation may stifle their curricular innovation, limit diversity and depth in teaching materials, and ultimately diminish the overall quality of education [60].

Third is decreased teamwork ability. Overreliance on GAI tools such as ChatGPT can weaken students’ communication skills and ability to engage actively in collaborative teamwork [72,152]. Furthermore, the frequent use of these tools limits opportunities for meaningful interpersonal interaction with peers and mentors, hindering the development of essential teamwork and communication skills [152].

Fourth is decreased practical problem-solving ability. Practical problem-solving is essential for clinical decision-making and patient management. However, the convenience of GAI tools may lead students to rely on preexisting solutions, neglecting the deeper analysis and logical reasoning necessary to develop personalized answers [52,55,74,75,77,87,151-153]. Furthermore, using these tools may reduce interaction with mentors and peers, limiting students’ opportunities to gain diverse perspectives through collaborative discussions and approach problems from multiple angles [147].

Ethical controversies can occur due to the following factors.

First is the authenticity of the test results, as the integration of GAI in testing and assessment may compromise the accuracy and effectiveness of traditional methods used to evaluate students’ actual capabilities. GAI-generated responses or GAI-assisted evaluations risk reflecting the performance of GAI itself rather than students’ authentic abilities. This issue is evident in various exams, such as medical licensing and specialty exams, presenting new ethical challenges in medical education [27,50,78,99,144,146,154].

Second is academic misconduct, since GAI-generated content often evades traditional plagiarism detection tools, making it easier for students to exploit GAI tools to complete assignments or write papers without being detected, thus jeopardizing academic integrity [31,154]. In addition, the ease of using GAI to generate answers cultivates students’ mindset of overreliance on such tools for academic tasks, which may increase their future likelihood of academic misconduct [70-72,124,155]. This issue extends beyond individual students and poses a broader threat to academic ethics, as GAI-generated content can be misinterpreted as original work, distorting academic evaluations [125,150].

Third is a lack of clinical interaction and emotional resonance. When addressing complex ethical or emotional medical issues, GAI lacks the empathy and emotional responsiveness inherent in human physicians, potentially undermining trust in the doctor-patient relationship [98,131]. This limitation is supported by a General Medicine In-Training Examination (GM-ITE) study comparing GPT-4 and Japanese residents. In the GM-ITE, “medical interview and professionalism” category assesses patient communication, ethics, and professionalism. It uses scenario-based questions (eg, addressing a terminally ill patient’s anxiety or resolving treatment ethics). Responses are scored 0‐10 based on communication appropriateness, empathy depth, and ethical application, with top marks for nuanced, human-centric judgment. Notably, GPT-4 scored 8.6 points lower here than residents [38]. Furthermore, because GAI tools do not provide an authentic, interactive experience or situational awareness, they may struggle to simulate the behavior and reactions of real patients accurately. This limitation makes it challenging for students to fully appreciate the importance of empathy and its application in doctor-patient interactions, which affects their development of communication and empathy skills development [38,54,76,77,101,107,147,152].

Fourth is resource inequality, which is most evident in the unequal access to technology and data. Datasets used for training GAI often exhibit biases, particularly involving data from different racial or socioeconomic backgrounds. This can worsen existing health care disparities. Furthermore, developing high-quality LLMs requires substantial computational resources, creating significant access barriers, especially for educational institutions or students with limited financial means. Hence, subscription fees and hardware limitations restrict their access to these GAI tools [67,74,85,134].

Fifth is the ownership of intellectual property rights. The widespread use of GAI in medical education raises numerous intellectual property concerns, particularly regarding copyright disputes related to the medical data used during AI training [113]. In addition, the legal status of GAI-generated content remains unclear, as current copyright laws do not adequately address the ownership of GAI-generated images and texts. This leaves the ownership of such content unclear, complicating the determination of whether the rights belong to the user, the developer, or other stakeholders [27,50,58,59,124].

Sixth is the “black box” problem and the attribution of responsibility. The application of GAI in medical education faces a significant challenge known as the “black box” problem. This issue arises from the lack of transparency and interpretability of GAI models, which directly affects the safety and reliability of these applications in medical settings. This lack of transparency makes it hard to understand how GAI reaches specific conclusions, especially when results are erroneous or biased, complicating efforts to trace and correct mistakes [86,88,148]. Furthermore, when GAI is used for diagnostic or clinical decision support, any errors or biases in its generated results can make it difficult to establish accountability. Trust in the doctor-patient relationship is built on clear responsibility. However, the lack of transparency in GAI models undermines this trust, leaving patients and physicians uncertain about the safety and reliability of GAI-driven decisions [36,114].

This study makes a key contribution to pedagogical practice by establishing the RMA tripartite framework and revealing its developmental imbalances, thereby providing a practical paradigm for the integration of GAI into medical education. The core value of this framework lies in elucidating the dynamic closed-loop nature of technology-enhanced education, wherein resource provision establishes the pedagogical foundation, methodological innovation activates knowledge transformation, and assessment feedback drives systemic evolution; these 3 components constitute an interlocking educational mechanism [165].

As evidenced in the results section, the current imbalance, characterized by rich exploration in GAI-supported educational resources and teaching methods yet relatively limited progress in GAI-driven automated evaluation of learner performance, stems from an overemphasis on short-term efficiency in early technology adoption. This has led to systemic neglect of assessment’s role as an optimization tool. For example, GAI is widely used to generate diverse clinical cases and pathological images to enrich educational resources and design adaptive learning pathways to innovate teaching methods. However, in learner assessment, most GAI tools still rely on simple automated scoring of knowledge-based quizzes, with few leveraging GAI to evaluate higher-order competencies such as clinical reasoning or diagnostic accuracy [26]. Another instance is that many researchers use GAI to create interactive simulation scenarios as a methodological advancement but fail to integrate automated assessment features that track learners’ decision-making processes in these scenarios [53]. This misses opportunities to use assessment data to refine the scenarios themselves. Overreliance on GAI for resource and method innovation without matching progress in automated learner assessment risks disconnecting what is taught or provided from what learners need to master, ultimately limiting GAI’s ability to drive meaningful change in medical education.

Achieving optimal integration requires establishing a bidirectional enhancement cycle centered on assessment. Automated assessment data capturing learning bottlenecks should guide the real-time expansion of clinical case libraries’ pathological spectra and difficulty calibration [166], shifting resource provision from one-size-fits-all to demand responsiveness. Simultaneously, the focus on core competencies (such as clinical reasoning and problem-solving) emphasized in teaching methods must be integrated into new assessment dimensions [167], driving teaching methods to evolve from mere knowledge transmission to competency development. Within this cycle, assessment functions not merely as a quality monitoring tool, but as the central nexus for the co-evolution of resources and methods.

Realizing this vision necessitates educators reconceptualizing operational logic [165]. This involves using assessment data to inform the development of educational resources, specifically leveraging insights into learners’ knowledge gaps and skill deficiencies to dynamically adjust the complexity of clinical cases [168], embedding real-time, practical, and contextual feedback mechanisms within high-order teaching activities like simulated diagnostics to optimize pedagogical strategies [169], and establishing adaptive rules enabling cross-dimensional interaction to facilitate systemic iteration [170,171]. Collectively, this structural transformation elevates the tripartite framework into an organic educational operating system.

However, technological integration inherently presents dual challenges, highlighting the importance of upholding core principles of human-AI collaboration. Generating educational resources without clinical context review risks reinforcing data biases [172]; methodological innovation overly reliant on algorithmic decisions may erode critical thinking [9]; and automated assessment replacing human judgment may overlook students’ psychological needs, reducing course engagement and well-being scores [173]. These manifestations of technological alienation arise from the partial ceding of human agency. Resolution lies in upholding a human-AI symbiotic vision: recognizing GAI as a collaborator, not a replacement, in educational evolution. Specifically at the resource layer, clinicians and educators must oversee the development of educational resources (eg, clinical cases) to balance efficiency, ethics, and clinical authenticity [174,175]. At the method layer, educators should direct learning path design to integrate technological augmentation with pedagogical wisdom [176]. At the assessment layer, institutions should implement verification systems that combine human evaluation with machine automation, ensuring assessments balance efficiency with humanistic dimensions [173,177]. This reconfiguration of responsibilities positions technology as a tool and reaffirms human stewardship of education.

This study identifies a technological imbalance in the application of GAI within medical education. This imbalance is characterized by the dominance of large general language models, while the development of specialized models for specific medical disciplines has lacked systematic progress. This limitation restricts the depth of technology-enabled education and indicates a neglect of multimodel collaborative mechanisms within current research paradigms.

The study proposes an integrated system using general LLMs alongside specialized medical models, employing a hierarchical collaborative architecture to reshape the technological ecosystem of medical education. The core operational logic establishes a 3-tiered functional division: general models act as the central hub for teaching interactions, handling basic task parsing and process orchestration; medical specialized models, drawing on vertical domain knowledge bases, execute high-complexity core teaching tasks such as clinical reasoning and medical image generation; and a cross-model validation mechanism forms a closed-loop quality control system. This architecture adapts the hospital’s multidisciplinary team approach to AI in education, aligning technological capabilities with the requirements of medical education for expertise, reliability, and contextual authenticity.

Within medical education, this integrated system can facilitate 3 key changes. First, it addresses limitations in specialized knowledge depth inherent in traditional general models, improving training efficacy for advanced clinical reasoning. Second, it leverages GAI’s multimodal capabilities, which integrate text and image data, to address key issues in medical imaging education including shortages of teaching resources like rare pathological images and the limits of static materials in showing dynamic anatomical relationships. This support helps evolve pathology visualization from static atlases to interactive 3D simulations, letting students explore spatial structures and pathological changes more intuitively [178]. Third, it establishes a cross-model knowledge validation chain to automatically identify and correct typical logical inconsistencies and factual errors in general models, ensuring the academic rigor of teaching content. These changes collectively represent a paradigm shift from tool-assisted learning to intelligent teaching partnership systems [179].

Supporting the effective operation of this system requires targeted solutions to key technical challenges. The primary task involves developing specialized models with medical context adaptive capabilities, specifically enhancing their semantic parsing of unstructured clinical texts to address performance variability in complex case analysis [180,181]. Concurrently, it is necessary to construct dynamically evolving medical education datasets that incorporate cross-regional case spectra and multilingual clinical literature to systematically mitigate cultural biases and time-lag effects in training data [182]. Integrating privacy-preserving computation techniques like federated learning can enable secure data collaboration among institutions, continuously optimizing model localization and adaptation while safeguarding patient information security [183,184].

This study reveals a pronounced regional disparity in the application of GAI within the field of medical education. Specifically, regions with a very high HDI dominate research output in this domain, while contributions from the low-HDI areas account for only 2%. The scarcity of cross-tier collaboration between very high- and low-HDI areas further exacerbates this structural inequity in resource distribution. This imbalance epitomizes systemic inequalities within global knowledge production systems, rooted in 3 compounding barriers: inadequate computational infrastructure in resource-constrained settings impedes technological localization, proprietary restrictions on core models under patent regimes limit feasible technology transfer, and excessive reliance on clinical data from high-income countries compromises model adaptability to regional health care priorities. Without deliberate intervention, this self-reinforcing Matthew Effect cycle risks intensifying the global fragmentation of medical educational resources [185].

Addressing this complex challenge necessitates a multitiered governance framework. At the international level, binding technology-sharing agreements should request that holders of advanced models provide architectural access under fair-use principles, emphasizing the need to balance innovation with equitable access, while emulating open-source paradigms as a reference model [186]. Concurrently, the World Health Organization could coordinate multinational efforts to develop nonprofit medical corpora incorporating disease spectra prevalent in low-HDI regions, such as tropical and endemic diseases [187]. Nationally, ministries of education should integrate computational infrastructure into public medical education budgets [188] and similar to the Medical Education Partnership Initiative (MEPI) [189], establish dedicated funds for cross-border institutional partnerships to co-develop localized pedagogical tools that address specific regional educational needs. Institutionally, medical schools should adopt algorithmic transparency protocols, requiring deployed GAI tools to provide auditable model documentation that details the demographics and geographical coverage of training data. Fairness assessments of these tools should be carried out by multidisciplinary committees, which include clinicians, ethicists, and community representatives [190].

Simultaneously, institutional responses must address secondary risks through integrated technical, educational, and regulatory safeguards. To counter academic misconduct, educational institutions should implement dual-track verification systems that require GAI-assisted submissions to be accompanied by generation logs and validated through detection tools [191]. Academic journals must establish clear authorship standards declaring proportional human-GAI contributions [192]. Mitigating critical thinking erosion requires curriculum committees to incorporate GAI-free clinical reasoning assessments, such as on-site case analyses evaluating independent diagnostic and management planning capabilities as prerequisites for professional certification [193].

Technical deficiencies demand targeted interventions. Reducing model hallucinations requires dynamic fact-checking systems linking GAI outputs to authoritative medical knowledge bases, with confidence levels displayed during teaching platform usage [194]. To address the opacity of algorithms, where the process by which GAI models derive conclusions remains unclear, it is necessary to document the diagnostic reasoning processes of these models. Such documentation allows instructors to review the reasoning, helps determine accountability when inconsistencies occur, and can be integrated into resident training evaluations to strengthen oversight of GAI-assisted decision-making [195].

Fundamentally, governance paradigms must transition from a technocentric approach to symbiotic development. Compared to the commonly used “human in the loop” [196], which mainly emphasizes humans overseeing or making final decisions in AI systems, symbiotic agency theory goes further: it highlights mutual shaping between humans and AI. Humans guide AI development through ethical norms and clinical experience, while AI enhances human capabilities by expanding cognitive boundaries, forming a dynamic, mutually reinforcing relationship [11]. Policies should affirm human primacy in medical education, exemplified by reserving clinical empathy training exclusively for human instructors while limiting GAI to standardized case supplementation. An effective return to the essence of symbiotic agency means building collaborative mechanisms as shown in Figure 5: educators lead in setting teaching goals and ensuring ethical alignment (eg, reviewing GAI-generated cases to match real clinical logic); GAI supports personalized learning; students provide feedback to refine GAI tools; and policies clarify rights and responsibilities in this interaction. This human-centered approach ensures technological advancement aligns with pedagogical integrity and global equity imperatives.