Thematic synthesis of the 42 studies [21-62] identified 7 major domains and 213 associated subtheme entries, each linked to representative quotations in Table S2 in Multimedia Appendix 4. The domains were as follows: (1) educational support and learning enhancement; (2) learning outcomes, educational effectiveness, and pedagogical value; (3) model performance, accuracy, and variability; (4) assessment, examination readiness, and question-type sensitivity; (5) reliability, safety, and risks; (6) adoption, usability, and implementation barriers; and (7) advanced AI architectures and future educational design.

Overall, the evidence was dominated by benchmarking and model performance evaluations of knowledge retrieval, question answering, and examination performance, with comparatively limited research on learner-centered outcomes, longitudinal learning, educational interventions, and real-world implementation. The 7 themes broadly encompassed educational applications and outcomes, technical performance and reliability, and implementation and future system design.

Across the included studies, generative AI was consistently portrayed as a flexible and scalable educational support resource that can supplement traditional ENT training by providing accessible, query-responsive support across different learning contexts and levels of training. Reported applications included facilitating professional discussion [27], bridging access to specialized knowledge [29], providing query-tailored explanations for early-stage learners [35], and supporting continuing medical education [46]. In this sense, AI is framed not as a replacement for educators but as a complementary educational tool alongside established learning resources and expert instruction [49,50].

A prominent dimension of educational support was the capacity of generative AI to retrieve, synthesize, and present knowledge efficiently. Advanced systems such as ENTAgents retrieved information from guidelines, textbooks, and academic literature within seconds, supporting efficient learning and revision [47]. ChatENT similarly streamlined access to specialty knowledge and reduced the need to manually search guidelines, textbooks, and academic papers [25]. In audiology, LLMs provided explicit reasoning processes that may help teachers and students rapidly identify key concepts and understand underlying reasoning processes [52]. Together, these findings suggest that the core educational contributions of generative AI lie in optimizing access to, organization of, and reasoning around knowledge rather than merely generating new content.

Beyond information access, several studies emphasized individualized learning support, formative feedback, self-assessment, and targeted remediation. LLMs such as GPT-4o provided immediate feedback on board-style questions and analyzed incorrect responses, helping learners identify specific knowledge gaps and pursue targeted remediation in a self-directed manner [48]. GPT-4 was proposed to generate personalized academic plans tailored to individual learning styles and abilities [29], while other models provided rapid feedback, explanations, and personalized remediation for examination preparation [61]. ChatGPT supported immediate self-assessment through past examination questions [60], while models with moderate-to-high board-question accuracy were proposed as study companions for basic knowledge reinforcement and practice-oriented learning [56]. These functions position generative AI as a mechanism for formative assessment, adaptive learning, self-assessment, and personalized remediation.

ChatGPT was reported to correctly interpret subtle domain-specific distinctions, including closely related anatomical concepts [27]. In pediatric otorhinolaryngology, AI was described as promising for theoretical learning and standardized clinical scenarios [55]. Beyond knowledge acquisition, emerging evidence suggests a potential role for generative AI in supporting scenario-based and reflective learning. Chatbots were used to explore clinical scenarios [22], while LLMs may function as scalable, case-sensitive reflective prompts rather than authoritative guides, encouraging trainees to reconsider operative plans and reflect on case-specific decisions [59]. Finally, emerging modalities extend educational support beyond text-based interaction. Multimodal systems such as Gemini integrated explanatory images and links into responses, providing additional study resources [53]. AI avatar–based educational platforms enabled rapid production of standardized educational content, reduced faculty workload, and facilitated swift revisions when guidelines changed [54]. Taken together, generative AI was positioned as a multifaceted educational adjunct that can broaden access to specialty knowledge, streamline information retrieval, support reasoning and scenario-based learning, provide personalized feedback and remediation, and extend educational delivery through multimodal and scalable formats. Importantly, its educational role is consistently framed as complementary to established learning resources and expert instruction rather than as a substitute for them [49,50,60].

The most direct evidence regarding learning outcomes comes from interventional studies embedding generative AI into structured, competence-adjusted educational systems. In the Multi-Expert Knowledge-Aggregated Adaptive Assessment Scheme (MEKAS) study, trainees receiving AI-assisted, competence-adjusted instruction demonstrated significantly greater improvements in clinical core medical knowledge in otolaryngology compared with those undergoing traditional education. The undergraduate- and postgraduate-year (UPGY) experimental group showed markedly higher score gains than controls, and similar benefits were observed in ENT residents, with no significant difference in educational effects between learner groups [41]. Beyond formal knowledge testing, an AI avatar–delivered ENT induction program significantly improved junior physicians’ self-reported confidence in areas including endoscopic anatomy recognition and clinical triage, although acceptability varied [54]. Together, these interventions demonstrated objective knowledge gains and improved learner confidence.

Educational effects were also observed in higher-order cognitive processing and clinical planning. In operative-planning scenarios, LLM-generated feedback altered residents’ clinical decision-making and surgical plans [59]. Rather than functioning as a decision substitute, the model served as a reflective aid, prompting learners to compare their plans with AI-generated alternatives [59]. This “reflection loop” suggests a pedagogical role beyond factual knowledge transfer, supporting structured reasoning, self-evaluation, and reflective practice. A second major line of evidence concerns the quality of AI-generated learning materials. LLMs, particularly GPT-4, generated ENT study guides with high accuracy, relevance, and clarity, closely aligned with residents’ curricular needs and judged by experts as legitimate educational resources [51]. ChatGPT was also considered potentially useful for problem-solving, critical thinking, and reading comprehension [43].

Importantly, pedagogical value was observed even where summative examination performance was limited. Although ChatGPT-3.5 would likely not have passed the American Board of Otolaryngology Written Question Exam, it provided well-constructed explanations and achieved performance comparable to the lower percentile of otolaryngology residents, suggesting educational benefit through explanation, clarification, and conceptual reinforcement [37]. Thus, educational value may extend beyond pass-fail performance to explanatory quality. AI-generated responses in theoretical and clinical scenarios could achieve quality scores comparable to human specialists [55], while ChatGPT-4o showed examination performance statistically comparable to that of ENT residents in 1 study [57].

Overall, generative AI showed potential to support knowledge acquisition, learner confidence, reflective reasoning, and high-quality educational content [37,41-43,51,54-57,59]. However, the current evidence remains stronger for short-term performance, confidence, and process-oriented outcomes than for sustained trainee-level educational benefit. Further studies are needed to determine whether these promising effects translate into durable gains in learning, clinical reasoning, and real-world training outcomes [40]. Current findings favor structured, pedagogically guided use over unguided standalone application.

Overall, the LLMs achieved moderate to high levels of accuracy on otolaryngology-related tasks, although performance varied substantially across datasets and evaluation settings. Reported accuracies ranged from modest levels, such as 53% on otolaryngology in-service practice examination questions, to high rates of comprehensive or correct responses in structured settings [23,26]. ChatGPT achieved passing-level performance across multiple conditions [24,33]; high accuracy and reproducibility were reported for salivary gland diseases [50]; and newer models performed particularly well in standardized knowledge assessments. For example, o3-mini and GPT-4o achieved accuracies of 93.35% and 90.45%, respectively, on otolaryngology United States Medical Licensing Examination–style questions, significantly exceeding the average medical student performance of 55.44% [61]. High answer-explanation concordance and validity were also reported, particularly for foundational anatomy and pathophysiology [24,25,33,60]. Taken together, current LLMs performed reasonably well on selected ENT knowledge tasks, while remaining strongly dependent on task structure and evaluation context.

A consistent observation across studies was the marked variability in performance across subspecialties, question categories, and difficulty levels [21,48]. LLMs performed better on easier items, whereas more nuanced and in-depth topics posed significant challenges [26]. Similarly, diagnostic tasks were characterized by satisfactory but not comprehensive performance, with models often producing correct answers without consistently providing complete or clinically robust reasoning pathways [31]. Models showed strong performance in theoretical knowledge areas such as cerebrospinal fluid otorrhea, hearing pathophysiology, temporal bone anatomy, and the vestibular system [57] but unsatisfactory performance on auditory physiology questions [22]. Multimodal capability was similarly variable: ChatGPT-4 demonstrated moderately acceptable image-recognition performance [43], while newer GPT-4 models showed substantial improvements on both text- and image-based questions [37]. Thus, accuracy should be interpreted in relation to task complexity, content domain, modality, and reasoning demands, rather than as evidence of generalized competence.

Cross-model comparisons demonstrated rapid generational improvement. Newer models, particularly GPT-4, consistently outperformed GPT-3.5 [37,48]. Standard and custom GPT-4 achieved accuracies of 72% and 81.3%, respectively, compared with approximately 51% to 53% for GPT-3.5 on comparable board-style questions [40], while another study reported 20% improvement in accuracy and 10% improvement in consistency with GPT-4 over GPT-3.5 [22]. In specific domains, such as anatomy learning and examination preparation, GPT-4o achieved particularly high accuracy, exceeding that of DeepSeek, Gemini, and Claude models [62]. However, rankings varied across evaluations: Gemini [48,53] and Bing Copilot [55] were top performers in individual benchmarks, Bard outperformed ChatGPT-3.5 overall and in otology [42], whereas ChatGPT-4 outperformed Bard in skull-base surgery [44]. These rankings reinforce that model superiority is dependent on the comparator, model generation, and evaluation task rather than representing a fixed hierarchy.

Model performance was also multidimensional. Gemini performed well in general otolaryngology question answering and conceptual clarification, whereas GPT-4 was effective for image-based questions and concise, high-accuracy explanations [53]. Claude 2 received the highest ratings for medical adequacy and conciseness, while ChatGPT-4 performed best in covering validated answers [28]. In audiology, ChatGPT generated the most complete responses, whereas Bing Chat performed best for concise relevance [22]. Thus, model performance is multidimensional, and model selection may need to consider the intended educational task and desired response characteristics—including accuracy, completeness, relevance, conciseness, and comprehensibility—rather than overall accuracy alone [22,28].

Variability was also evident in the consistency and stability of model outputs. ChatGPT-4 showed greater consistency than ChatGPT-3.5 [22]. ChatGPT and DeepSeek likewise demonstrated substantial response consistency despite minor fluctuations in an audiologist qualification examination [52], while high reproducibility was reported for ChatGPT responses concerning salivary gland diseases [50]. ChatGPT-5 led a guideline-anchored evaluation in accuracy, comprehensiveness, guideline adherence, and temporal stability [58].

Performance was sensitive to linguistic and structural features of questions. GPT-4 models demonstrated significantly better handling of negated questions than nonnegated ones, whereas Claude 3 Haiku performed poorly on negated questions [48]. Comparable concordance and validity across short-answer and other structured formats further indicated a complex interaction between performance and question structure [25].

Comparisons with human learners and clinicians revealed a more nuanced performance gradient. Earlier evaluations found that LLMs remained inferior to human expertise [21,31], although some were competitive with junior or mid-level residents [35,38]. Conversely, ChatGPT-4 outperformed residents across training years in 1 study, although the performance gap narrowed with increasing residency experience [45], while ChatGPT-4o outperformed residents on selected theoretical knowledge items [57]. Newer models such as o3-mini and GPT-4o also significantly exceeded average medical student performance on standardized knowledge questions [61]. Relative performance of LLMs versus human learners depends substantially on trainee experience, model generation, and the type of competence being assessed; strong performance on knowledge-based questions should therefore not be interpreted as equivalent to specialist-level clinical expertise.

Taken together, contemporary LLMs showed improving accuracy, validity, consistency, and reproducibility, with newer models increasingly reaching or exceeding human learner performance on selected standardized knowledge tasks. However, performance remains heterogeneous across model generations, platforms, subspecialties, modalities, question structures, and reasoning demands. Notably, an accuracy plateau of approximately 73% to 76% among leading general-purpose models has also been reported in specialized otolaryngology knowledge assessment, suggesting persistent limitations without domain-specific adaptation [56]. Thus, high benchmark performance should be interpreted as task-specific capability rather than evidence of uniform competence across the breadth and complexity of otolaryngology [21,24,26,31,37,48,53,56,57,62].

Across the included studies, the examination performance of LLMs varied substantially according to model generation, assessment format, and task demands, with evidence ranging from suboptimal accuracy to passing or high performance on selected professional examinations. Accuracy in single-answer multiple-choice questions (MCQs) was often 40% to 60% [27], and ChatGPT answered 51% of otolaryngology board-preparation questions correctly [39]. However, advanced models demonstrated stronger performance: GPT-4 passed selected sleep medicine certification questions [29], ChatGPT-4 achieved 75% accuracy and passed the Taiwan Audiologist Qualification Examination [43], and both ChatGPT and DeepSeek exceeded the 60% passing threshold on a standardized audiologist qualification examination [52]. More recent models showed further gains, with o3-mini maintaining high accuracy across difficulty levels and outperforming students, particularly on more challenging questions [61]. Thus, examination readiness has improved with newer models but remains model dependent and task dependent.

A particularly robust and reproducible pattern across studies is the pronounced sensitivity of LLM performance to question format. Multiple evaluations demonstrated significantly poorer performance on MCQs compared with single-choice or open-ended formats. In 1 study, ChatGPT achieved correct responses in more than 60% of single-choice questions but only approximately one-third of MCQs [21]. Multiple-choice underperformance was also statistically significant across all tested models (P<.001), suggesting that this limitation reflects a general structural weakness rather than a model-specific deficiency [48]. However, later studies indicate that multiple-choice performance is not uniformly poor. ChatENT reduced errors on MCQs by 26% relative to ChatGPT and outperformed ChatGPT-4 on both multiple-choice and short-answer board questions [25], while DeepSeek and ChatGPT achieved near-passing or passing performance on complex MCQs and high accuracy on reverse-logic questions [52]. Guideline-constrained multiple-choice tasks also produced tightly clustered performance across newer models, suggesting that structured questions anchored to established clinical pathways may reduce response drift [58]. Thus, question-type sensitivity appears to interact with model architecture and task structure rather than representing a fixed limitation of all LLMs.

Question difficulty represents another major determinant of assessment performance. A domain-specific ENT GPT assistant demonstrated a pronounced decline in accuracy from 86.8% on the easiest questions to 42.9% on the most difficult questions [46], whereas o3-mini maintained high performance across all difficulty levels, including 100% accuracy at the greatest difficulty level in 1 evaluation [61]. These findings suggest that sensitivity to difficulty remains model dependent and may diminish as model capabilities advance. Assessment performance was also strongly influenced by modality. GPT-4V performed better on text-only than image-based board examination questions, although providing both textual and visual information improved accuracy for image-based items [36]. GPT-4 performed poorly on direct image interpretation and frequently appeared to infer answers from contextual information rather than the images themselves [45], while Bard was unable to process more than half of image-based questions because of file-processing errors [42]. Graph-based questions were also below the passing threshold [52]. Although newer models showed improvement [42], text-based performance may therefore overestimate competence when visual interpretation is required.

Beyond format and modality, factual knowledge performance did not necessarily reflect higher-order clinical reasoning. Board-style questions may predominantly assess factual recall [38]. ChatGPT-4o performed strongly on theoretical knowledge questions but less well on clinical decision-making, contextual evaluation, and patient-specific judgment [57]. Similarly, general-purpose LLMs performed better on rule-based practice-management content than on clinical subspecialty questions requiring more nuanced reasoning [56]. Narrative case prompts also remained more demanding than guideline-constrained questions, particularly with respect to completeness, stepwise management, and calibration [58]. Thus, strong knowledge-based examination performance should not be equated with clinical reasoning or judgment.

This distinction also has implications for assessment design. Several studies proposed that future examinations should place greater emphasis on higher-order competencies, including diagnostic reasoning, clinical judgment, and image interpretation, rather than relying predominantly on factual knowledge questions that current LLMs increasingly answer successfully [38,45]. ChatGPT-generated otorhinolaryngology MCQs produced correct and unambiguous answer keys in 70% of cases and showed high topic relevance and focus; however, vignette and distractor quality were weaker [34]. AI-assisted question generation therefore remains promising, but expert review remains necessary before formal assessments.

Overall, newer LLMs increasingly achieved passing or high performance on selected examinations but remained sensitive to question format, difficulty, modality, and cognitive demands [36,38,40,42,45,46,50,52,56-58]. They may support examination preparation and formative assessment, but benchmark performance may overestimate real-world competence, particularly when assessments emphasize constrained knowledge tasks [56]. Their improving performance also strengthens the rationale for assessments that explicitly evaluate higher-order reasoning and clinically meaningful competencies.

A central concern across the included studies was the reliability of generative AI in medical education and clinical knowledge support. Although some evaluations reported high response quality in terms of clarity, relevance, and usefulness, these apparent strengths were consistently undermined by poor source transparency and unreliable citation practices [49,55]. Chatbots produced incorrectly attributed or spurious citations [49], while other studies reported absent source attribution [22] and unreliable or nonexistent references in open-ended responses [32]. Accordingly, successful educational integration requires accurate answers alongside transparent reasoning, reliable sourcing, alignment with curricular goals, and safeguards against outdated or fabricated guidance [60,61].

A major limitation is hallucination. One study reported hallucinated content in more than 13% of open-ended responses, despite apparently coherent and confident outputs [27]. Models also hallucinated during sequential image-based questions, producing similar but incorrect answers across different questions [49]. Additional studies showed that LLMs may produce human-like responses without the contextual reasoning of experienced clinicians [28], may be vulnerable to incorrect or insufficient input information [43], and may generate biased or overly simplified answers when complex medical problems are reduced to fact-based representations [42]. Such errors may be particularly difficult for inexperienced learners to recognize [42].

Safe educational use requires users to retain sufficient biomedical knowledge and critical appraisal skills to recognize inaccurate or oversimplified outputs. One study specifically warned of intellectual dependency and reduced perceived need to learn complex pathophysiology [40]. This concern is particularly relevant for inexperienced learners, supporting the need for expert supervision to prevent uncritical acceptance of confidently presented inaccuracies [26,42,59]. Beyond educational reliability, errors also posed potential patient safety risks. In 1 example, ChatGPT provided incorrect anatomical information regarding surgical planes for an osteoplastic flap, which could compromise outcomes if applied clinically [33]. Several authors cautioned against the routine use of generative AI for clinical decision-making or for unsupervised educational purposes, particularly in contexts where outputs may influence patient care [23,26]. ChatENT reduced safety-related concerns relative to more general-purpose systems [25], suggesting that knowledge grounding and system design can improve—but not eliminate—safety limitations. Residual errors therefore continue to require educator judgment and oversight when generative AI is incorporated into teaching [52].

Reliability concerns extended to AI-generated educational materials. ChatGPT-generated MCQs were generally grammatically correct but sometimes used medical terminology inaccurately or imprecisely, weakening their reliability for formal assessment [34]. The overall quality of such questions was judged to require thorough expert review and revision before use in examinations [34]. These findings reinforce that linguistic fluency should not be equated with domain accuracy and that AI-generated educational content requires specialist validation before curricular or assessment use. Temporal consistency was another reliability dimension. In a 1-month retest of next-generation LLMs, the leading systems preserved their relative performance ranking and demonstrated limited output drift [58]. Thus, reliability should be conceptualized not only in terms of whether an answer is correct but also whether model behavior remains stable, reproducible, and appropriately aligned with established guidance over time.

Generative AI should not be relied on as a stand-alone authority, especially in high-stakes contexts. Multiple studies explicitly advised against overreliance on AI chatbots for final interpretations, assessments, or clinical judgments, emphasizing that such systems should function only as adjuncts to traditional educational resources and expert supervision [23,49]. Even in controlled educational settings, some studies reported that although peer review of AI-generated scripts can mitigate accuracy risks, the lack of interactivity and emotional or pedagogical nuance still limits educational effectiveness and learner acceptance [54]. This adjunctive model is reinforced by newer studies recommending controlled educational use rather than autonomous clinical decision-making [56,57]. In operative-planning education, a supervised “dual-plan” approach—in which residents first generate an independent plan and subsequently compare it with an LLM-generated alternative—was proposed as a way to preserve reflective and metacognitive benefits while reducing the risk of adopting incorrect suggestions uncritically [59]. These findings favor structured workflows that preserve independent reasoning, human judgment, and expert verification.

Overall, current generative AI systems remain constrained by hallucination, unreliable or absent source attribution, susceptibility to flawed inputs, imprecise terminology, residual error, and limitations in contextual and clinical reasoning. Additional risks include intellectual dependency and the uncritical internalization of confidently presented inaccuracies, particularly among inexperienced learners. Although domain-specific grounding and newer model generations may improve safety and temporal stability, they do not remove the need for expert validation and educator oversight. Accordingly, the available evidence supports the use of generative AI as a supervised and transparently governed educational adjunct rather than as an independent source of medical knowledge or clinical decision-making authority [22,25-28,32,34,38,40,42,43,49,52,55-61].

A recurrent facilitator of adoption was ease of use, which influenced users’ willingness to engage with AI-based educational systems. In the MEKAS intervention, both ENT residents and UPGY trainees reported high perceived ease of use scores, supporting usability as an important determinant of acceptance [41]. Similarly, ENTAgents avoided the need for model fine-tuning by using external knowledge bases and self-reflective prompting, thereby lowering technical barriers to deployment [47]. High reproducibility across otolaryngology topics may further support user trust and implementation [23].

Despite these technical advantages, actual adoption among trainees remains limited. Survey data from Italian ENT residents showed modest ChatGPT use, with only a small proportion of trainees having incorporated it into their academic or scientific activities, despite broad awareness of its potential functions [30]. Implementation may also be constrained by unequal institutional access to advanced models and infrastructure, potentially widening disparities in training opportunities [59]. Current use was concentrated on research-related and low-risk academic support tasks rather than core studying. Trainees reported higher perceived usefulness for activities such as citation searching, content searching, and language revision, whereas studying and examination preparation received lower ratings [30]. Generative AI therefore appears to function more as a productivity and academic support tool than as a central learning instrument, possibly reflecting uncertainty about its pedagogical reliability. Larger multicenter studies incorporating diverse assessment formats and updated AI models have been called for to establish safe and effective strategies for integrating generative AI into medical education [57].

Overall, usability, low technical barriers, and reproducibility support implementation feasibility but do not ensure adoption. Wider pedagogical integration requires equitable access, stronger evidence of educational effectiveness across diverse settings, and clearly defined strategies for safe implementation [23,30,41,47,57,59].

A prominent direction in the literature was the shift from general-purpose chatbots toward knowledge-grounded and explainable educational systems. The MEKAS platform exemplifies this approach through a knowledge-based AI design that explicitly prioritizes both accuracy and explainability, supporting learning of clinical core medical knowledge in otolaryngology [41]. This reflects a broader shift toward adapting LLMs to domain-specific knowledge and reasoning requirements rather than relying solely on model scale.

Another trend was the integration of retrieval-augmented generation (RAG) and agent-based frameworks. ENTAgents used agents and RAG to supplement LLMs with up-to-date, specific information from curated knowledge bases and external sources, improving accuracy and completeness [47]. The system also incorporated self-correction and reflective reasoning to refine responses and improve factual grounding [47]. Domain-specific customization also improved specialty performance. ChatENT outperformed ChatGPT-4 on both multiple-choice and short-answer board questions [25], while custom GPT-4 models achieved higher accuracy than standard versions [40]. The ENT GPT assistant demonstrated that customization to specialty-specific language, terminology, and clinical nuances improved contextual relevance and accuracy [46]. Together, these findings favor combining knowledge grounding with specialty customization and expert validation rather than relying on general-purpose models alone [25,28,40,42,46]. In multimodal education, improved medical-image performance may require specialized prompting, advanced image-recognition technologies, and training specifically focused on medical images [36]. Domain refinement, real-time learning mechanisms, and improved interpretability have also been proposed [38].

Generative AI may also support pedagogical designs involving automated generation of case scenarios, differential diagnosis exercises, interactive learning modules, standardized patient simulations, and assessment questions and rubrics, all of which can support scalable, consistent, and resource-efficient instructional design across teaching sessions [49]. Customizable synthetic cases may further expand these applications [29], while virtual-patient scenarios may allow residents and students to practice diagnostic and therapeutic skills in a safe simulation environment [35]. AI has also been proposed for integration into examination simulations that test basic knowledge recall and allow residents to assess gaps in foundational knowledge [45]. In this context, AI is framed not merely as a content generator but as an infrastructure for supporting structured learning workflows, including review, concept reinforcement, board-style question practice, and rapid knowledge recall [49]. Its ability to interact with learners, provide explanations, and respond to user customization also creates opportunities for more interactive and adaptive educational experiences [40].

These systems may further promote active learning and higher-order cognitive skills. AI-generated assessment questions can be used to encourage active engagement rather than passive text review, a strategy that aligns with established principles of memory retention and learning science [49]. At the same time, instructors are encouraged to guide learners to critically appraise AI-generated responses and identify reasoning errors, supporting analytical reasoning, reflective judgment, and diagnostic rigor [49]. In operative-planning tasks, residents were reported to engage with LLM outputs reflectively rather than passively, particularly when AI-generated recommendations differed from their own plans [59]. Future frameworks may therefore promote metacognition by requiring learners to formulate, compare, critique, and revise their reasoning rather than passively accept AI-generated answers [59], while supervised guideline-linked systems may support case discussion and formative assessment [58]. At the architectural level, the literature further suggests that future progress may require explicit support for medical reasoning rather than continued increases in general model scale alone. A reported performance plateau among general-purpose LLMs suggests that adaptation to medical reasoning patterns may represent a more important next step [56]. Hybrid systems combining LLMs with structured knowledge bases or clinical decision trees may provide explicit reasoning frameworks for complex cases [56].

Overall, the evidence suggests a transition from general-purpose conversational AI toward hybrid, domain-adapted, and pedagogically governed educational systems integrating knowledge grounding, retrieval augmentation, specialty customization, structured reasoning, simulation, adaptive interaction, and reflective learning [25,28,29,35,36,38,40-42,45,46,56,58,59]. These approaches position generative AI less as an autonomous answer generator and more as a component of explainable, adaptive, and educationally aligned personalized learning platforms.