Metaverse and generative AI in education: a scoping review of technologies, applications, evaluation, challenges, and future directions

1 Introduction

The convergence of Artificial Intelligence (AI), Generative AI (GenAI), and Extended Reality (XR) technologies has contributed to growing research interest in the Metaverse as an environment for digital interaction, content creation, and immersive learning (Chen et al., 2024; Chamola et al., 2024; Basyoni et al., 2025; Rahimi et al., 2025). The Metaverse generally comprises persistent, interactive virtual environments in which users engage through avatars and spatial interfaces (Al-Ghaili et al., 2022; Tukur et al., 2024; Stephenson, 2003). In education, existing literature has explored the use of Virtual Reality (VR), Augmented Reality (AR), Mixed Reality (MR), and AI-driven content generation to support immersive learning environments, personalized instruction, intelligent tutoring systems, conversational agents, and adaptive educational experiences (Wang and Huang, 2025; Mittal et al., 2024; Darejeh and Sepasgozar, 2026; Sharma et al., 2025; Juarez et al., 2025b). However, the extent to which these technological capabilities translate into sustained educational benefits remains an open empirical question.

Despite growing research activity, AI-enabled Metaverse education continues to face important technical and methodological challenges. Existing studies report that many systems remain prototype-oriented, with limited empirical validation and restricted evidence of real-world deployment. In addition, scalability, interoperability, AI trustworthiness, privacy, and evaluation rigor have been identified in the literature as continuing challenges to the maturity and adoption of these educational systems (Chen et al., 2024; Basyoni et al., 2025; Tabassum et al., 2025; Chu et al., 2025; Hui et al., 2025; Lan and Azimi, 2025; Hussain et al., 2024; Wong et al., 2024).

Beyond technological development, assessing the educational value of AI-enabled Metaverse environments also requires a pedagogical perspective. Established learning theories–including constructivism, experiential learning, cognitive load theory, embodied cognition, social constructivism, and self-regulated learning–provide complementary lenses for interpreting potential pedagogical mechanisms associated with learner engagement, knowledge construction, collaboration, and adaptive instruction (Marougkas et al., 2023; Yan et al., 2025; Makransky and Petersen, 2021; Hsu and Lee, 2026; Zhang et al., 2025). Accordingly, this review examines AI-enabled Metaverse systems as both technological and pedagogical environments, while distinguishing proposed educational potential from empirically demonstrated outcomes.

Existing research varies considerably in scope and analytical emphasis across enabling technologies, educational applications, pedagogical foundations, technical implementation, user evaluation, governance, and future research directions. As summarized in Table 1, several existing reviews provide valuable multidimensional perspectives; however, their coverage of these dimensions remains uneven, and no identified review systematically integrates all of them within a single analytical framework.

TABLE 1. Comparison of existing reviews with this study.

✓ = substantive coverage; Partial/Limited/Moderate/Minimal = varying degrees of coverage; –/None = not substantively addressed; Tech. Impl. and Lim. = Technical Implementation and Limitations; User Eval. = user evaluations; Future Dir. = future directions; SD = ScienceDirect; WoS = web of science.

To complement these existing efforts, this study conducts a scoping review of Metaverse and GenAI-enabled applications in education. Specifically, it develops an integrative, multidimensional taxonomy and synthesis encompassing Metaverse technologies, GenAI approaches, educational application domains, pedagogical foundations, technical implementation and limitations, user evaluation practices, governance considerations, and future research directions, providing a consolidated evidence base for this rapidly evolving field.

The study is guided by the following research questions:

• RQ1: What are the taxonomies of Metaverse-enabling technologies and environments, GenAI technologies, and their applications in education? Rationale: AI, GenAI, XR, and supporting infrastructures constitute the technological backbone of Metaverse systems. Systematically identifying and categorizing these technologies clarifies their roles, interdependencies, and integration patterns. Furthermore, existing studies span diverse educational contexts, including virtual classrooms, skill training, simulation-based learning, and collaborative environments. A structured taxonomy is therefore required to organize these applications and understand their alignment with educational objectives.
• RQ2: What are the major technical and methodological limitations in current AI-enabled Metaverse education systems? Rationale: Identifying recurring technical and methodological challenges provides insight into current bottlenecks and informs the development of more robust and reliable educational systems.
• RQ3: What evaluation practices and user-centered studies are employed in existing works? Rationale: Understanding how learning effectiveness, usability, and user experience are evaluated is essential for assessing the maturity and real-world readiness of AI-enabled Metaverse education systems.
• RQ4: What future research directions can advance AI-driven Metaverse education? Rationale: Synthesizing emerging trends and proposed research directions helps identify priorities for advancing future Metaverse-enabled educational systems.

Accordingly, this review provides the following contributions:

• It provides an integrative, multidimensional taxonomy that hierarchically organizes Metaverse-enabling technologies and environments according to immersive modalities, persistent/system environments, and computational representations, alongside GenAI approaches, educational applications, and supporting infrastructures.
• It critically synthesizes pedagogical foundations, technical limitations, and user-centered evaluation practices, revealing a key maturity gap between the technological capabilities reported in existing studies and the comparatively limited evidence of sustained educational outcomes, real-world effectiveness, accessibility, inclusivity, and AI trustworthiness.
• It identifies evidence gaps and summarizes future research directions for Metaverse and GenAI-enabled education.

Collectively, these contributions provide an integrated evidence-based understanding of Metaverse and GenAI-enabled education by connecting reported technological capabilities, educational applications, pedagogical perspectives, evaluation practices, and research gaps.

The remainder of this paper is organized as follows. Section 2 reviews the related work. Section 3 describes the research methodology. Section 4 presents the application domains and discusses the taxonomy of enabling technologies and other contextual factors. Section 5 examines technical and methodological limitations, while Section 6 analyzes user perceptual studies. Future research directions are outlined in Section 7. Section 8 discusses ethical, governance, and sustainability considerations. Section 9 critically discusses the principal findings through the cross-cutting tensions emerging from the reviewed evidence and summarizes the strengths and limitations of the review. Finally, Section 10 concludes the paper.

2 Related work

Recent years have witnessed a growing number of secondary studies, including surveys, systematic reviews, scoping reviews, and bibliometric analyses, examining the convergence of Artificial Intelligence (AI), Generative AI (GenAI), Extended Reality (XR), and Metaverse technologies, particularly in educational contexts. These studies provide valuable synthesized insights into technological capabilities, educational applications, implementation challenges, and emerging research priorities. However, they differ considerably in scope, methodological approach, temporal coverage, and the extent to which they address technological, pedagogical, evaluative, governance, and future-oriented dimensions.

Several reviews focus specifically on AI-enabled Metaverse systems for education. Chen et al. (2024) provide a conceptual and analytical overview of AI-enabled Metaverse environments, highlighting their potential to support personalized learning, automated content generation, adaptive assessment, intelligent tutoring, and immersive educational experiences. The study also discusses enabling technologies and challenges related to technological accessibility, infrastructure, data privacy, pedagogical integration, and ethical considerations. However, the review remains largely conceptual and provides limited systematic synthesis of empirical user evaluation practices and technology adoption trends across implemented educational systems.

Similarly, Almeman et al. (2025) present a systematic literature review examining the integration of AI and Metaverse technologies in education. Their review provides substantial coverage of immersive technologies, AI-enabled educational applications, technical implementation issues, and user-centered evaluation, while also considering pedagogical and governance-related concerns. However, its primary emphasis remains on the broader integration of AI and Metaverse technologies, and it does not differentiate emerging GenAI subcategories, such as proprietary versus open-source foundation models. Furthermore, because its coverage ends in 2024, it does not capture more recent developments in multimodal GenAI, conversational tutors, and increasingly sophisticated AI-driven educational assistants.

From a GenAI perspective, Mittal et al. (2024) provide a comprehensive review of GenAI applications in education. Their study examines how generative models support personalized learning, educational content creation, intelligent tutoring, assessment, simulations, and conversational interfaces, while also discussing challenges related to privacy, bias, security, explainability, and accountability. The review additionally considers the potential convergence of GenAI and the Metaverse for immersive education. Nonetheless, its primary emphasis remains on GenAI for education, with comparatively limited synthesis of technical implementation and empirical user evaluation within immersive and Metaverse-based learning environments.

Extending this line of work, Qian et al. (2023) systematically review Metaverse applications, paradigms, and emerging technologies in education and introduce the concept of Meta-AI, emphasizing the potential synergy between GenAI technologies, including Large Language Models (LLMs), and immersive learning environments. The review provides broad coverage of educational applications, enabling technologies, pedagogical approaches, and user evaluation practices. However, its literature coverage ends in 2022 and therefore predates many recent developments in multimodal GenAI, AI agents, and more advanced GenAI-enabled educational systems.

A number of survey papers focus more directly on the role of GenAI in enabling the Metaverse. Chamola et al. (2024) provide a comprehensive survey of GenAI for the Metaverse, categorizing generative models according to the type of content they produce, including text, images, video, audio, 3D assets, and gaming content. Their work highlights the role of GenAI in large-scale content creation, immersive interaction, personalization, and virtual environment generation. However, the primary emphasis is on Metaverse infrastructure and content-generation capabilities, with comparatively limited attention to educational applications and governance, and without systematic synthesis of pedagogical foundations or learner-centered evaluations.

Basyoni et al. (2025) complement this perspective by reviewing GenAI-driven Metaverse systems through the lens of AI-Generated Content (AIGC), discussing technological opportunities alongside robustness, adversarial risks, human-centered considerations, governance, and responsible innovation. While their analysis provides valuable technological and governance insights, education constitutes only one application domain, and educational deployment and learner-centered evaluation receive comparatively limited attention.

Similarly, Lv (2023) presents a broad overview of GenAI technologies in the Metaverse, emphasizing their role in accelerating content production, supporting multimodal generation, and enhancing creative and interactive capabilities. The review discusses several technical architectures and limitations associated with generative models, but adopts a broad technological perspective and does not substantively address educational applications, pedagogical foundations, or empirical user evaluation.

Beyond technological capabilities, some secondary studies concentrate on governance, ethical, and legal concerns. Tabassum et al. (2025) conduct a scoping review of ethical and legal challenges associated with GenAI and the Metaverse, identifying concerns related to bias, misinformation, privacy, accountability, transparency, intellectual property, and regulatory governance. Although the review provides important societal and regulatory insights, its emphasis is on ethical and legal governance rather than educational applications, pedagogical foundations, technical implementation, or empirical user evaluation.

From an immersive systems perspective, Rahimi et al. (2025) survey the integration of GenAI with Virtual Reality under the concept of Generative VR. Their review examines multimodal content generation, adaptive virtual environments, intelligent agents, real-time interaction, technical architectures, system limitations, governance considerations, and future research directions. However, the review remains primarily VR-centric, with education representing only one application area and with comparatively limited synthesis of educational outcomes, pedagogical foundations, and learner-centered evaluation practices.

Nawaz et al. (2025) provide a bibliometric analysis of Metaverse research in education, identifying dominant technologies such as VR and AR and emerging themes related to personalized learning, constructivist and experiential learning, immersive teaching, AI-supported education, and scalable learning environments. This study offers a valuable macro-level understanding of the evolution of Metaverse-based education research, including pedagogical and governance-related themes. However, as a bibliometric analysis, its primary emphasis is on publication and thematic trends rather than detailed synthesis of technological implementations and empirical user evaluation practices.

More recent secondary studies published in 2025–2026 further illustrate the growing convergence of AI, GenAI, immersive technologies, and education. Wen et al. (2026) systematically review 41 empirical studies on AI-powered immersive learning environments and provide substantial coverage of educational contexts, pedagogical considerations, technical implementation, methodological limitations, and user evaluation. Their findings indicate that many interventions remain small-scale, short-term, and laboratory-based, with limited theoretical grounding and inconsistent use of validated evaluation instruments. Although GenAI is identified as an emerging trend, the review adopts a broader AI perspective in which GenAI represents only one subset of the technologies considered.

Similarly, Salah et al. (2026) review AI integration, security, privacy, and governance in educational Metaverse environments. Their study provides broad coverage of educational applications, pedagogical foundations, technical implementation, security mechanisms, privacy-preserving approaches, and governance, while proposing a dedicated framework for secure and ethical AI integration. However, the review is primarily oriented toward security, privacy, and trustworthy AI, and GenAI receives comparatively less focused treatment than broader AI technologies.

In a broader XR context, Ning et al. (2026) systematically review the integration of GenAI with XR, examining VR, AR, MR, multimodal interaction, GenAI architectures, user evaluation, governance, technical limitations, and future directions. Although education and training are represented among the identified application domains, the review spans multiple sectors and is not specifically centered on educational implementations or pedagogical foundations.

Complementing these perspectives, Zhong et al. (2026) survey GenAI ethics literacy in the Metaverse and propose a knowledge–attitude–competence framework to characterize responsible GenAI use. Their review provides substantial coverage of education, pedagogical considerations, ethical risks, privacy, transparency, bias, misinformation, governance, and future research needs. However, technical implementation and empirical user evaluation receive comparatively less systematic attention because the primary focus is GenAI ethics literacy rather than comprehensive characterization of immersive educational systems.

Collectively, these reviews demonstrate the growing maturity of research at the intersection of Metaverse technologies, GenAI, AI, and education. Existing syntheses provide substantial coverage of individual and, in several cases, multiple dimensions, including educational applications, immersive technologies, GenAI, pedagogy, technical implementation, user evaluation, security and privacy, ethical governance, and future directions. However, as summarized in Table 1, the analytical scope and emphasis vary considerably across studies, and no identified review systematically integrates all of these dimensions within a single evidence-based framework.

2.1 Positioning of the present study

As summarized in Table 1, prior reviews provide valuable and, in several cases, multidimensional perspectives on AI-enabled Metaverse education. Nevertheless, their coverage differs across the key analytical dimensions of Metaverse technologies, GenAI, education, pedagogy, technical implementation and limitations, user evaluation, governance, and future research directions. The present study complements these efforts by jointly synthesizing these dimensions within a single scoping review framework.

Specifically, this review synthesizes evidence from 114 primary studies published or indexed from January 2022 up to the respective database search dates in 2026 across five major scholarly databases. It integrates enabling Metaverse technologies, GenAI approaches, educational application domains, pedagogical foundations, technical implementation and limitations, user evaluation practices, governance considerations, and future research directions. By operationalizing these dimensions within a unified analytical framework, the review provides a consolidated evidence base for understanding the current state, remaining gaps, and emerging directions of Metaverse and GenAI-enabled education.

3 Methodology

To address the research questions outlined in Section 1, this study adopts the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) guidelines (Tricco et al., 2018). A scoping review approach was selected because the field of Metaverse and GenAI-enabled education is broad, rapidly evolving, and characterized by heterogeneous study designs, technologies, application domains, and evaluation methods. Unlike systematic reviews or meta-analyses, which typically focus on narrowly defined questions and quantitative effect synthesis, this study aims to comprehensively map, categorize, and synthesize the existing body of knowledge, identify research gaps, and highlight emerging trends across diverse types of evidence.

PRISMA-ScR is a widely recognized framework that ensures transparency, methodological rigor, and comprehensive reporting in scoping reviews, particularly when synthesizing evidence from diverse study designs and domains. As illustrated in Figure 1, the framework establishes a structured workflow that supports a systematic and reproducible review process. A completed PRISMA-ScR checklist, specifying the location of each reported item within the manuscript, is included in the Supplementary Material. Although the review protocol was not prospectively registered (e.g., with the Open Science Framework), the review methodology, eligibility criteria, search strategy, and data extraction procedures were established before the literature search commenced and remained unchanged throughout the review. The study was conducted and reported in accordance with the PRISMA-ScR guidelines to ensure methodological rigor, transparency, and comprehensive reporting. Furthermore, to promote transparency and reproducibility, the complete review protocol, search strategies, coding framework, extracted dataset, and replication materials are publicly available in the project’s replication repository and Supplementary Material.

3.1 Search strategy

The search strategy was structured around four key components: the selection of relevant bibliographic databases, the formulation of search terms and query strings, the definition of inclusion and exclusion criteria, and the implementation of a systematic study selection process.

Note: The initial electronic database searches were conducted in February 2026, covering studies published or indexed from January 2022 up to the respective search dates. Following the expansion of the IEEE Xplore and ACM Digital Library searches from primarily title-based retrieval to abstract-level retrieval, these two databases were re-searched in June 2026. Accordingly, studies available and meeting the eligibility criteria up to the respective database search dates were considered for inclusion. The exact search dates, search fields, query strings, filters, and retrieval counts for each database are provided in the Supplementary Material (Digital_Libr_Info_Full).

3.2 Data source selection

To ensure comprehensive coverage of the relevant literature, five major academic databases were systematically queried: ACM Digital Library, IEEE Xplore, Scopus, ScienceDirect, and Springer. These databases were selected due to their extensive indexing of peer-reviewed research across Artificial Intelligence, Metaverse technologies, GenAI, and technology-enhanced education. Collectively, they capture both the underlying technological foundations (e.g., Large Language Models, Extended Reality, and immersive systems) and the pedagogical dimensions (e.g., learning outcomes, user engagement, and adaptive learning) that are central to AI-driven and Metaverse-enabled educational systems. This selection ensures strong domain relevance while maintaining broad and multidisciplinary literature coverage.

3.3 Search terms

The following terms were identified as representative of the core concepts used to construct the search queries:

• Metaverse terms: Metaverse, virtual worlds, immersive virtual environments, Extended Reality, XR, Virtual Reality, VR, Mixed Reality, MR, Augmented Reality, AR, Digital Twins.
• Generative AI terms: Generative AI, Generative Artificial Intelligence, Large Language Models, LLM, AI-driven, AI, Artificial Intelligence, Machine Learning, ML, Deep Learning, DL.
• Education terms: education, learning, teaching, training, professional education, vocational training, lifelong learning.

3.4 Search strings

Search queries were developed using a combination of keyword variants and semantic synonyms to ensure comprehensive retrieval of relevant studies. Table 2 presents the search strings used across the selected data sources ¹ .

TABLE 2. Restricted electronic search results for Metaverse, GenAI, and education.

3.5 Search criteria

Studies were selected based on their direct relevance to educational applications leveraging Metaverse and GenAI technologies. The inclusion and exclusion criteria were defined as follows:

3.5.1 Inclusion criteria

• I1: The study provides direct evidence relevant to one or more of the research questions.
• I2: The study explicitly addresses the integration of Metaverse and GenAI in transforming education ² .
• I3: The publication is a peer-reviewed journal article, conference paper, workshop paper, congress, or symposium contribution.
• I4: The study was published or indexed from January 2022 up to the applicable database search date (February 2026 for the initial searches and June 2026 for the updated IEEE Xplore and ACM Digital Library searches).

The year 2022 marks the mainstream adoption of GenAI, following the public release of systems such as DALL⋅ E 2, Stable Diffusion, and ChatGPT, which enabled widespread content creation.

• I5: The study is written in English.
• I6: The paper exceeds three pages to ensure sufficient depth of discussion and analysis.

3.5.2 Exclusion criteria

• E1: Studies that do not address any of the research questions were excluded.
• E2: Studies that do not explicitly address the integration of Metaverse and GenAI in education (e.g., purely conceptual or speculative works).
• E3: Non-peer-reviewed publications, surveys, and secondary studies were excluded.
• E4: Studies published before 2022 were excluded.
• E5: Non-English publications were excluded.
• E6: Duplicate records, papers with fewer than four pages, and studies without accessible full texts were excluded.

Borderline cases were handled conservatively to maintain the conceptual scope of the review. Studies were excluded when the integration between Metaverse/XR and GenAI was only implicit and could not be clearly established; when one technology was merely incidental or mentioned without an identifiable role in the educational system; or when the AI component represented generic or conventional AI rather than a technology identifiable as belonging to the GenAI family. Conversely, studies that did not explicitly use the term “GenAI” were retained when the reported AI method or model could be clearly identified as generative and was meaningfully integrated with the immersive educational application.

3.6 Study selection

The study selection process was carried out in three stages: (1) initial identification of records, (2) removal of duplicates followed by title and abstract screening for relevance, and (3) full-text evaluation to determine final eligibility based on the criteria outlined in Section 3.5.

3.6.1 Initial identification phase

This phase represents the first stage of the review process, during which the predefined search terms (see Section 3.3) were applied across all selected databases. As illustrated in Part A of Figure 1, this step captures the total number of records retrieved using broad search queries ( n = 5,436 ). Following database-specific filtering strategies, including publication date range, document type, language, and metadata field restrictions, 2,553 records were excluded, leaving a refined subset of 2,883 records. This subset is reported as the “Restricted Electronic Search Results” in Table 2. This refinement step improves the relevance and manageability of the retrieved records prior to subsequent screening.

3.6.2 Screening based on duplicate removal, title, and abstract

In this stage, duplicate entries were first identified and removed, resulting in the exclusion of 84 records. Following automated duplicate removal, records retrieved from multiple databases were manually cross-checked during the title/abstract and full-text screening stages to identify overlapping publications reporting the same study. Where overlap was identified, only a single, most complete record was retained for data extraction and synthesis. The remaining 2,799 records were then screened based on their titles and abstracts, alongside the availability of full-text versions. A further 2,431 records were excluded due to lack of relevance to the review scope or unavailability of full texts, leaving 368 articles for full-text eligibility assessment.

3.6.3 Full-text screening

The shortlisted articles were then subjected to a detailed full-text review in accordance with the inclusion and exclusion criteria defined in Section 3.5. Studies were excluded if they only made superficial references to education (or related concepts) without substantive contribution, or if they did not explicitly involve Metaverse and GenAI technologies. Following this rigorous assessment, 254 full-text articles were excluded, resulting in a final set of 114 studies identified as relevant and included as primary sources.

The multidisciplinary nature of the review is evident in the diversity of publication venues (see Supplementary Table S1). Figure 1B illustrates the distribution of publication types, while Figure 1C presents the temporal distribution of the included studies over recent years. Supplementary Table S1 further summarizes the distribution of the selected studies across journals, conferences, symposia, workshops, and congresses, highlighting the broad dissemination of research on Metaverse and GenAI-enabled education.

To enhance methodological rigor and reduce selection bias, the screening and study selection process was conducted independently by three co-authors. Each full-text article was reviewed independently by at least two reviewers. Studies for which both reviewers reached the same decision were accepted directly, whereas disagreements were resolved through discussion and consensus with a third (senior) co-author.

As an additional methodological validation, a post hoc inter-rater reliability assessment was conducted after completion of the screening process. An independent reviewer re-screened a randomly selected sample of 110 articles from the 368 full-text studies assessed for eligibility (approximately 30%), using the original inclusion and exclusion criteria without reference to the final consensus decisions.

Agreement between the original screening decisions and the independent reassessment was quantified using Cohen’s kappa ( κ = 0.834 ), indicating almost perfect agreement according to the Landis and Koch interpretation Landis and Koch (1977). Although the inter-rater reliability assessment was conducted post hoc rather than prospectively, the high Cohen’s κ (0.834) supports the consistency of the full-text eligibility decisions. Nevertheless, because this reliability assessment was retrospective, it should be interpreted as an additional methodological validation rather than as a substitute for prospectively planned inter-rater reliability assessment.

Microsoft Excel was used to support screening and data extraction, while Overleaf’s .bib workflow facilitated reference management, ensuring consistency and collaborative tracking throughout the review process. To strengthen pedagogical interpretation, the extraction process also recorded whether each study explicitly or implicitly aligned with major learning-theory constructs, including constructivist/experiential learning, cognitive-load management, embodied interaction, social or collaborative learning, and self-regulated learning. Additional details, including the replication package consisting of the list of randomly selected studies used for the post hoc inter-rater reliability assessment, are provided in Supplementary Material.

3.7 Data extraction and coding

Following the final study selection phase, a structured data extraction and coding procedure was conducted to systematically analyze the included studies. A standardized data extraction framework was developed using Microsoft Excel to ensure consistent collection and organization of relevant study characteristics across all selected papers. The extracted information included publication metadata (e.g., title, year, publisher, venue, and publication type), application domains, Metaverse technologies, GenAI categories, degree of freedom (DoF), target user groups, technology maturity (envisioned versus implemented systems), implementation maturity (prototype versus validated systems), user evaluation characteristics, first-author affiliation countries, technical limitations, and future research directions.

For variables permitting multiple classifications, a multi-label coding approach was employed. Categories were non-mutually exclusive; therefore, an individual study could be assigned to multiple application domains, Metaverse technology categories, GenAI categories, target user groups, technical limitation categories, and user evaluation categories, where applicable. Consequently, the cumulative category counts for these variables may exceed the total number of included studies ( n = 114 ), and the corresponding percentages may sum to more than 100%. Percentages for each category were calculated independently using the total number of included studies ( n = 114 ) as the denominator.

The coding process was conducted iteratively using a combination of open coding and thematic synthesis. Initially, recurring concepts, technologies, pedagogical themes, evaluation characteristics, and technical limitations were identified through open coding of the selected studies. These preliminary codes were subsequently refined and grouped into higher-level thematic categories through multiple rounds of independent review, discussion, and consensus-building.

To improve coding consistency and reproducibility, operational definitions and decision rules were established for the principal analytical variables. Application domains were coded according to the substantive educational context, discipline, or training purpose addressed by each study; because educational contexts may overlap, multi-label assignments were permitted when more than one domain was substantively represented. Metaverse technologies were coded when an immersive or virtual-environment technology (e.g., VR, AR, MR, XR, Metaverse platforms, Digital Twins, or Virtual Campuses) was explicitly reported as part of the educational system or experience. GenAI was coded when the study explicitly identified, or sufficiently described, a generative model or technique capable of generating new content, including text, images, audio, video, 3D content, or other educational resources. Generic AI methods without an identifiable generative function were not coded as GenAI. An AI-enabled system was defined as an educational system in which AI performed an identifiable functional role, such as content generation, tutoring, feedback, personalization, assessment, interaction, or adaptation.

Implementation maturity distinguished envisioned or conceptual systems from implemented prototypes and empirically validated systems. User perceptual studies were coded when end users participated in an evaluation involving measures of their perceptions, experiences, learning, usability, engagement, acceptance, or related outcomes. Pedagogical foundations were coded when a study explicitly referenced or demonstrated substantive alignment with an established learning theory or pedagogical construct. Technical limitations were coded when a study reported technological, implementation, deployment, performance, reliability, data, integration, scalability, or related constraints affecting the proposed system or approach. Table 3 summarizes these operational definitions and corresponding coding rules.

TABLE 3. Operational definitions and decision rules for the principal coding variables.

This iterative process enabled the development of the taxonomies and analytical dimensions presented in Section 4 and the subsequent synthesis sections of the manuscript. To enhance methodological rigor and minimize selection and interpretation bias, the screening, coding, and data extraction activities were performed independently by multiple reviewers. Each full-text article was reviewed by at least two co-authors, while disagreements regarding study eligibility, categorization, or extracted information were resolved through discussion and consensus, with a third senior reviewer consulted when necessary.

Consistent with the objectives of scoping reviews and the PRISMA-ScR guidelines, a formal methodological quality or risk-of-bias assessment was not undertaken because the primary aim was to comprehensively map and synthesize the breadth of the existing literature rather than evaluate intervention effectiveness.

Instead, the review systematically characterized the maturity of the available evidence through multiple complementary dimensions, including technology maturity (envisioned versus implemented systems), implementation maturity (prototype versus validated systems), evaluation methodologies, validation approaches, DoF, and user study characteristics. These complementary indicators provide important context for interpreting the overall maturity and robustness of the current evidence base while remaining consistent with the objectives of a scoping review.

To enhance transparency and reproducibility, the complete coding schema, data extraction framework, search records, inter-rater reliability sample, and other replication materials are provided in the Supplementary Material (Replication File), including the Coding/Data_Extraction_Process, IncludedStudies_Metadata, Digital_Libr_Info, and other supporting analysis sheets.

4 Taxonomy of Metaverse and generative AI-based technologies for education (RQ1)

This section presents an integrative, multidimensional taxonomy of Metaverse- and GenAI-enabled educational technologies derived from the synthesis of 114 selected studies. The taxonomy is structured across six key dimensions: application domains, Metaverse technology types, GenAI categories, DoF, target users, and geographic distribution of first authors’ affiliations. These dimensions collectively provide a multi-perspective understanding of the technological, pedagogical, and contextual landscape of emerging AI-driven immersive education systems.

4.1 Application-oriented taxonomy

The reviewed studies span a diverse set of application domains, as illustrated in Table 4 and Figure 2c. To provide a deeper understanding of the application landscape, this subsection first defines each of the nine identified domains summarized in Table 4, followed by an analytical interpretation of their distribution.

• STEM Education: This domain focuses on science, technology, engineering, and mathematics learning through immersive simulations, virtual laboratories, and AI-assisted problem-solving. GenAI supports adaptive tutoring, automated feedback, and inquiry-based experimentation in safe and scalable virtual environments. Representative reference(s): Song et al. (2024), Luo et al. (2024), Li et al. (2024), Lan and Azimi (2025), Kryvoshyya et al. (2025).
• Higher Education: This domain includes undergraduate and postgraduate learning environments that utilize virtual campuses, immersive classrooms, and collaborative spaces. GenAI enhances these systems through personalized tutoring, automated content generation, and intelligent academic support tools. Representative reference(s): Chu et al. (2025), Abdelmagid et al. (2025), Hemminki-Reijonen et al. (2025), Yoo et al. (2025), Pang et al. (2025).
• Healthcare and Medical Education: This domain focuses on clinical training, procedural practice, and patient interaction through immersive simulations such as virtual anatomy exploration and surgical training. GenAI enables adaptive case generation, conversational virtual patients, and real-time feedback for competency-based learning. Representative reference(s): Koike and Majima (2025), Chheang et al. (2024), Vadisetty et al. (2025), Kadri et al. (2025), Miyawaki et al. (2024).
• Social Sciences and Humanities: This domain includes subjects such as history, sociology, philosophy, and cultural studies, emphasizing contextual understanding and experiential learning. Metaverse environments and GenAI are commonly used for historical reconstruction, immersive storytelling, and interactive cultural experiences. Representative reference(s): Juarez et al. (2025a), Juarez et al. (2025b), Hui et al. (2025), Wong et al. (2024), Zhu et al. (2025).
• K-12 and Special Education: This domain emphasizes accessibility, engagement, and personalized learning for younger learners and students with diverse educational needs. Applications commonly use gamified immersive environments and AI-driven adaptive learning pathways to support foundational and social-emotional skills. Representative reference(s): Cheng et al. (2024), Lyu et al. (2024), Padmavathi et al. (2024), Chen et al. (2025), Cunha et al. (2024).
• Language and Communication Education: This domain focuses on language acquisition, communication skills, and conversational practice through immersive interaction with AI-driven agents. GenAI supports real-time feedback, contextual dialogue generation, pronunciation correction, and multilingual learning. Representative reference(s): Ding and Song (2025), Zhong Y. et al. (2025), Pan et al. (2025), Viju and Terence (2025), Cheng et al. (2025).
• Safety and Risk Training: This domain prepares learners for hazardous or high-risk environments through realistic immersive simulations. GenAI enhances these systems with adaptive scenarios, context-aware guidance, and dynamic feedback to support situational awareness and decision-making. Representative reference(s): Nolter and Doherty (2025), Manolopoulos et al. (2025), Naderi and Shojaei (2025), Taylor et al. (2025), Hussain et al. (2024).
• Vocational and Workforce Training: This domain focuses on practical and job-specific skill development through hands-on virtual training environments. GenAI provides intelligent guidance, automated assessment, and adaptive instruction to support upskilling and workforce preparation. Representative reference(s): Nofal et al. (2025), An et al. (2025), Wang et al. (2026), Ji et al. (2025).
• Business and Entrepreneurship Education: This domain supports the development of business, innovation, and entrepreneurial skills through virtual marketplaces, collaborative simulations, and decision-making scenarios. GenAI enhances these environments by generating business cases, market scenarios, and strategic recommendations. Representative reference(s): Abdelmagid et al. (2025), Yang et al. (2025), LE CORRE and HUANG (2024).

TABLE 4. Distribution of main application domains across reviewed studies (n = 114).

The detailed study identifiers (S1–S114) corresponding to each application domain are provided in Supplementary Table S2 to improve the readability of the main manuscript. Categories are non-mutually exclusive; individual studies may be assigned to multiple categories. Therefore, cumulative counts and percentages may exceed n = 114 and 100%, respectively.

4.1.1 Analysis of application domains

The distribution of application domains (see Table 4) provides important insights into where Metaverse and GenAI technologies are currently being adopted in education. STEM education constitutes the largest category, with 27 studies (24%), followed closely by higher education with 26 studies (23%). The prominence of these two domains reflects the strong alignment of immersive environments and GenAI with technically intensive and academically structured learning contexts. Across both domains, the literature consistently emphasizes immersive visualization, virtual laboratories, simulation-based experimentation, intelligent tutoring, and AI-assisted content generation. Their substantial representation also reflects the greater availability of digital infrastructure, technical expertise, and research investment within universities and STEM-oriented institutions. Moreover, studies in these domains generally demonstrate comparatively stronger empirical validation through controlled experiments, usability evaluations, and learner performance assessments.

Healthcare and medical education represents the third largest application area, accounting for 17 studies (15%). In contrast to STEM and higher education, which primarily focus on conceptual understanding and collaborative learning, healthcare applications concentrate on procedural competence, clinical decision-making, and risk-free skills training. The combination of immersive simulations with GenAI-powered virtual patients and adaptive clinical scenarios enables realistic practice while minimizing training risks. This growing body of work illustrates the suitability of AI-enhanced immersive environments for disciplines that require repetitive practice, immediate feedback, and high-fidelity experiential learning.

Moderate research attention is observed in social sciences and humanities (14 studies, 12%), K-12 and special education (11 studies, 10%), and language and communication education (11 studies, 10%). These domains broaden the educational scope beyond technically oriented disciplines by emphasizing learner engagement, accessibility, inclusivity, cultural experiences, immersive storytelling, adaptive instruction, and conversational AI. Language education, in particular, leverages LLMs to support contextual dialogue, language practice, and personalized feedback, while K-12 and special education frequently explore adaptive learning pathways that accommodate diverse learner needs. Although these application areas continue to expand, many studies remain exploratory and comparatively fewer report large-scale or longitudinal evaluations.

Vocational and workforce training contributes 10 studies (9%), highlighting the growing use of immersive technologies for professional skills development, workplace simulation, and competency-based learning. Despite its practical importance, research activity in this area remains noticeably lower than in academic settings, suggesting that challenges related to infrastructure requirements, domain-specific customization, organizational readiness, and deployment at scale continue to limit broader adoption.

Safety and risk training accounts for 7 studies (6%), where immersive simulations and AI-generated dynamic scenarios are primarily employed to support emergency response, hazard recognition, and situational decision-making. These applications benefit from realistic environments that allow learners to safely experience high-risk situations while receiving adaptive feedback. In comparison, business and entrepreneurship education remains the least represented category, with only 3 studies (3%), despite the considerable potential of Metaverse environments for entrepreneurial training, collaborative problem-solving, negotiation, and experiential business simulations.

Taken together, the application landscape demonstrates a clear correspondence between educational objectives and the capabilities of Metaverse and GenAI technologies. Technically intensive disciplines predominantly exploit immersive simulations, analytical reasoning, and adaptive instructional support, whereas learner-centered domains place greater emphasis on personalization, accessibility, communication, and contextual engagement. The current evidence also reveals an uneven distribution of research effort, with STEM, higher education, and healthcare receiving substantially greater attention than vocational, business, and several socially oriented disciplines. Broadening empirical investigations across these underrepresented domains presents an important opportunity to improve the scalability, inclusiveness, and cross-disciplinary applicability of AI-driven immersive learning systems.

From a pedagogical perspective, the observed distribution of application domains reflects the complementary roles of established learning theories in shaping Metaverse- and GenAI-enabled education. Constructivism underpins the strong emphasis on immersive, experiential, and authentic learning environments that enable learners to actively construct knowledge through interaction and problem-solving. Social constructivism is evident in collaborative virtual environments that facilitate peer learning, communication, and knowledge co-construction, particularly in higher education and collaborative learning contexts.

Cognitive Load Theory highlights the importance of designing immersive interfaces and AI-assisted instructional support that minimize extraneous cognitive load while optimizing learning efficiency, especially in technically demanding domains such as STEM and healthcare education. Likewise, Self-Regulated Learning is increasingly supported through GenAI-driven personalized feedback, adaptive scaffolding, intelligent tutoring, and learning analytics, enabling learners to monitor, regulate, and improve their own learning processes across diverse educational contexts. Collectively, these findings suggest that the educational value of Metaverse and GenAI technologies depends not only on technological sophistication but also on their alignment with established pedagogical principles.

4.2 Metaverse technology taxonomy

To improve conceptual precision, the identified Metaverse-related technologies are organized hierarchically according to their functional and conceptual roles rather than treated as equivalent technology types. As summarized in Table 5 and Figure 2a, the taxonomy comprises three layers: (1) Immersive Modality, including VR, AR, MR, and XR, with XR treated as an umbrella category encompassing multiple immersive modalities; (2) Persistent/System Environment, comprising Metaverse platforms and virtual campuses; and (3) Computational Representation, represented by Digital Twins. These layers are non-mutually exclusive because an educational system may simultaneously employ an immersive modality, operate within a persistent virtual environment, and incorporate a computational representation.

• Layer 1 — Immersive Modality
• Virtual Reality (VR): VR refers to fully immersive computer-generated environments accessed through head-mounted displays and interactive tracking systems (Tukur et al., 2025). In education, VR is widely used for simulation-based learning, virtual laboratories, engineering design, and clinical training due to its strong support for experiential and spatial learning.
• Representative reference(s): Nofal et al. (2025), Juarez et al. (2025a), Song et al. (2024), Juarez et al. (2025b), Koike and Majima (2025).
• Augmented Reality (AR): AR overlays digital content onto the physical world through smartphones, tablets, or wearable devices. In education, AR supports contextual and interactive learning by enhancing physical environments with visual instructions, 3D models, and real-time information.
• Representative reference(s): Cheng et al. (2024), Lyu et al. (2024), Viju and Terence (2025), Villegas-Ch et al. (2024), Miyawaki et al. (2024).
• Mixed Reality (MR): MR combines physical and digital environments by enabling virtual objects to interact dynamically with the real world. Educational MR applications support highly interactive and hands-on learning experiences that integrate real and simulated components.
• Representative reference(s): Garzarella et al. (2025), Vaze et al. (2024), Barrón-Estrada et al. (2026), Cunha et al. (2024), Gutiérrez Maquilón et al. (2024).
• Extended Reality (XR): XR is an umbrella category encompassing VR, AR, and MR technologies rather than a modality at the same conceptual level. In education, XR commonly describes integrated immersive systems that support flexible deployment across different devices and levels of immersion.
• Representative reference(s): Zhong Y. et al. (2025), Kryvoshyya et al. (2025), Vadisetty et al. (2025), Cunha et al. (2024), Gianni et al. (2025).
• 2. Layer 2 — Persistent/System Environment
• Metaverse: Metaverse platforms represent persistent and shared virtual ecosystems integrating immersive technologies, social interaction, avatars, and collaborative environments. Educational Metaverse platforms commonly support virtual classrooms, collaborative learning spaces, and user-generated educational content.
• Representative reference(s): Chu et al. (2025), Abdelmagid et al. (2025), Hui et al. (2025), Ding and Song (2025), Yang et al. (2025).
• Virtual Campus: Virtual campuses are persistent educational environments that digitally represent or extend institutional spaces such as classrooms, laboratories, libraries, and social areas. They support collaborative, remote, and hybrid learning through immersive interaction and real-time communication.
• Representative reference(s): Juarez et al. (2025a), Juarez et al. (2025b).
• 3. Layer 3 — Computational Representation
• Digital Twin: Digital Twins are dynamic computational representations of physical systems or environments that may be continuously updated using real-time data. In education, they support simulation, monitoring, scenario testing, and data-driven decision-making.
• Representative reference(s): Sakkas et al. (2025), Taylor et al. (2025), Xu et al. (2025).

TABLE 5. Hierarchical distribution of Metaverse-enabling technologies and environments across reviewed studies ( n = 114 ).

Categories are non-mutually exclusive; individual studies may be assigned across multiple layers and categories. Consequently, cumulative counts and percentages may exceed n = 114 and 100%, respectively. *XR is treated as an umbrella category encompassing multiple immersive modalities rather than as a modality conceptually equivalent to VR, AR, or MR. Detailed study identifiers (S1–S114) corresponding to each category are provided in Supplementary Table S3.

4.2.1 Analysis of the Metaverse technology taxonomy

The hierarchical distribution of Metaverse-enabling technologies and environments, summarized in Table 5 and further detailed in Supplementary Tables S7, S8, reveals distinct patterns across the three conceptual layers.

At the immersive-modality layer, Virtual Reality (VR) is by far the most widely represented modality, appearing in 70 studies (61%), followed by Augmented Reality (AR) in 28 studies (25%) and Mixed Reality (MR) in 15 studies (13%). The high frequency of VR-based studies suggests that VR currently occupies a dominant position in the reviewed evidence base. It is particularly prominent in STEM, healthcare, and professional training, where the reviewed applications commonly involve procedural practice, complex visualization, hands-on experimentation, and immersive simulation. This high representation may partly relate to the comparatively mature ecosystem of VR hardware, software platforms, and educational applications.

AR enriches real-world environments with digital content, making it well suited for contextual learning, laboratory guidance, engineering visualization, and field-based instruction. Its ability to integrate virtual information into physical settings enables learners to maintain situational awareness while benefiting from interactive educational support. MR similarly integrates physical and digital components but enables more dynamic interaction between them, supporting collaborative and hands-on learning experiences.

XR is reported in 12 studies (11%) and is treated in this taxonomy as an umbrella category encompassing multiple immersive modalities rather than as a modality conceptually equivalent to VR, AR, or MR. XR-based approaches commonly integrate immersive capabilities to provide flexible learning experiences across different devices, interaction configurations, and levels of immersion.

At the persistent/system-environment layer, Metaverse platforms appear in 15 studies (13%), whereas virtual campuses appear in 3 studies (3%). Unlike immersive modalities, these categories characterize broader persistent and socially connected educational environments. Metaverse platforms primarily emphasize persistent virtual worlds, avatar-based interaction, collaborative learning, and shared educational ecosystems. Virtual campuses extend this concept toward institution-oriented digital environments integrating classrooms, laboratories, libraries, and social spaces. Their comparatively limited representation may reflect the greater infrastructural, interoperability, and organizational requirements associated with persistent, large-scale educational environments.

At the computational-representation layer, Digital Twins appear in 3 studies (3%). Rather than constituting an immersive modality or persistent environment, Digital Twins provide computational representations of physical entities, systems, or environments that support simulation, monitoring, scenario testing, and intelligent decision-making. Their educational implementation often requires real-time data integration, sophisticated computational resources, and supporting infrastructure, which may partly explain their limited representation in the reviewed literature.

Importantly, these three layers represent complementary dimensions rather than mutually exclusive alternatives. For example, an educational Metaverse platform may employ VR as its immersive modality while simultaneously incorporating a Digital Twin as its computational representation. Accordingly, individual studies may be coded across multiple layers where applicable.

Technology selection across the reviewed studies also shows patterns associated with the pedagogical objectives of individual application domains. Simulation-intensive disciplines predominantly employ VR, which may be related to its capacity to support immersive and realistic simulation, whereas AR and MR are more frequently adopted in applications involving interaction with the physical environment. Persistent Metaverse environments and virtual campuses increasingly support collaborative and socially connected learning experiences, while Digital Twins enable data-driven simulation and interaction with representations of physical systems.

Differences are also evident in evaluation maturity across the taxonomy. VR-based systems commonly report controlled experiments, usability studies, and learner performance assessments, reflecting the comparatively mature state of VR research in education. Conversely, Digital Twins and virtual campuses are more frequently represented by prototypes or early-stage implementations with limited empirical validation. Expanding rigorous evaluations of these emerging approaches represents an important direction for future research, particularly as educational institutions seek scalable, data-driven, persistent, and interoperable immersive learning environments.

4.3 Evidence-derived classification of generative AI technologies and ecosystems

This subsection presents an evidence-derived classification of the GenAI technologies and ecosystems reported across the reviewed studies, as summarized in Table 6 and Figure 2b. Consistent with the objectives of this scoping review, the classification reflects how GenAI technologies are identified and reported in the primary studies rather than imposing a general technical ontology of GenAI architectures. Accordingly, the identified categories may represent provider or deployment ecosystems, model families or generative approaches, open-source/model-distribution ecosystems, or cases in which the specific LLM was not reported. These distinctions are explicitly indicated below.

• OpenAI Ecosystem (GPT, DALL ⋅ E, Whisper): This category represents the OpenAI ecosystem and includes proprietary models used for conversational tutoring, adaptive feedback, automated content generation, speech recognition, and multimodal interaction in immersive educational systems. Their widespread representation in the reviewed studies may partly relate to their API accessibility and broad general-purpose capabilities. Representative reference(s): Chu et al. (2025), Abdelmagid et al. (2025), Nofal et al. (2025), Song et al. (2024), Juarez et al. (2025b).
• Hugging Face/Open-Source Model Ecosystem: This category captures studies reporting open-source models or model-distribution ecosystems such as Hugging Face, including transformer-based models such as GPT-Neo, BERT, and T5. These approaches support customizable and locally deployable AI applications and are used for personalized tutoring, domain-specific adaptation, and privacy-sensitive educational deployments. Representative reference(s): Viju and Terence (2025), Manolopoulos et al. (2025), Vijaya et al. (2025), Imam et al. (2025), Gianni et al. (2025).
• Diffusion/Generative Media Model Family: This category captures diffusion-based and related generative-media approaches, including systems such as Stable Diffusion and MidJourney. These models support immersive content creation, virtual scene generation, instructional visualization, and creative design applications within Metaverse-based education. Representative reference(s): Hui et al. (2025), Ding and Song (2025), Kryvoshyya et al. (2025), Garzarella et al. (2025), Constantinescu et al. (2024).
• Google AI Ecosystem (Gemini, Flan, Cloud AI): This category represents Google’s GenAI ecosystem, including cloud-based services and multimodal models. These systems are used for scalable tutoring, automated educational content generation, multimodal interaction, and cloud-native immersive learning applications. Representative reference(s): Cheng et al. (2025), Padmavathi et al. (2024), Varshita et al. (2025), Vijaya et al. (2025), Barrón-Estrada et al. (2026).
• Meta AI Ecosystem (LLaMA Family): This category represents Meta’s GenAI ecosystem, particularly the LLaMA family of open-weight LLMs. These models are integrated into immersive educational systems for conversational agents, adaptive tutoring, and AI-driven teaching assistants, particularly in VR/XR environments. Representative reference(s): Hui et al. (2025), Wang et al. (2026), Cheng et al. (2025), Alipour et al. (2025), El-Hajji et al. (2025).
• Chinese LLM Ecosystem (e.g., DeepSeek, ChatGLM, Hunyuan): This category captures Chinese LLM ecosystems reported across the reviewed studies, including DeepSeek, ChatGLM, and Hunyuan. These models support multilingual interaction, localized educational content, conversational tutoring, and regionally adapted immersive learning applications. Representative reference(s): Song et al. (2024), Hui et al. (2025), Luo et al. (2024), An et al. (2025), Barrón-Estrada et al. (2026).
• Unspecified/General LLM Reporting: This category captures studies that report the use of an LLM or general GenAI capability without providing sufficient information to identify the specific model or provider. It is retained as a reporting-status category to transparently represent the available evidence rather than being interpreted as a distinct model family or deployment ecosystem. Representative reference(s): Lan and Azimi (2025), Pang et al. (2025), Zhong Y. et al. (2025), Nolter and Doherty (2025), Zhong X. et al. (2025).
• GAN-Based Model Family: This category captures Generative Adversarial Network (GAN)-based approaches, primarily used for image synthesis and visual content generation. Although comparatively less common in the reviewed literature, GANs continue to appear in specialized immersive educational simulations and media-generation applications. Representative reference(s): Vadisetty et al. (2025), Sakkas et al. (2025).

TABLE 6. Evidence-derived distribution of reported generative AI technologies and ecosystems across reviewed studies ( n = 114 ).

This evidence-derived classification reflects how GenAI technologies are identified and reported in the reviewed studies and is not intended as an ontology of technically equivalent model classes. Categories may represent provider/deployment ecosystems, model families or generative approaches, open-source/model-distribution ecosystems, or reporting status. “Unspecified/General LLM Reporting” denotes studies that do not provide sufficient information to identify the specific model or provider. Categories are non-mutually exclusive; individual studies may be assigned to multiple categories. Consequently, cumulative counts and percentages may exceed n = 114 and 100%, respectively. Detailed study identifiers (S1–S114) are provided in Supplementary Table S4.

4.3.1 Analysis of reported generative AI technologies and ecosystems

Insights from Table 6, Supplementary Table S8, and Figure 2b illustrate the distribution of GenAI technologies and ecosystems as reported across the reviewed studies. Because the classification is evidence-derived, the categories capture different characteristics of the reported implementations, including provider/deployment ecosystems, model families or generative approaches, open-source/model-distribution ecosystems, and reporting status. They should therefore be interpreted as descriptive categories of the available evidence rather than as technically equivalent GenAI model classes.

The OpenAI ecosystem (GPT, Whisper, and DALL⋅E) constitutes the most frequently reported category, appearing in 68 studies (60%). Its widespread use highlights the prominent role of OpenAI technologies in enabling conversational tutoring, adaptive feedback, multimodal content generation, intelligent assistance, and natural language interaction within immersive learning environments. Their popularity is further supported by mature APIs, strong generalization capabilities, extensive developer ecosystems, and integration flexibility, allowing researchers to rapidly prototype and deploy AI-enhanced learning experiences.

Unspecified or general LLM reporting occurs in 31 studies (27%). These studies describe the use of LLM capabilities without providing sufficient information to identify the underlying model or provider. This category therefore represents reporting status rather than a distinct model family or ecosystem. Its relatively high frequency also highlights an important transparency and reproducibility issue, as incomplete model reporting can make it difficult to compare implementations or reproduce GenAI-enabled educational systems.

Hugging Face/open-source model ecosystems appear in 13 studies (11%). Their adoption reflects increasing interest in customizable and potentially locally deployable AI infrastructures that can provide greater flexibility and institutional control. Such approaches are particularly relevant to educational settings requiring domain-specific adaptation, reduced dependency on proprietary APIs, or greater control over data and deployment configurations.

Diffusion and generative-media model families are reported in 12 studies (11%), demonstrating the growing importance of visual and multimodal content creation within immersive learning environments. Rather than emphasizing conversational interaction, these approaches primarily support image synthesis, virtual asset generation, creative design, and immersive scene construction. Their integration complements language-based GenAI by enriching Metaverse environments with dynamically generated visual content.

Google AI and Meta AI ecosystems each appear in 9 studies (8%). Although their representation remains considerably lower than that of the OpenAI ecosystem, both contribute to the diversification of GenAI adoption in education. Google’s ecosystem is commonly associated with cloud-based and multimodal AI services, whereas Meta’s LLaMA family provides open-weight alternatives that offer flexibility for customized implementations and local deployment. Their adoption demonstrates increasing interest in alternatives that balance performance, accessibility, and deployment flexibility.

Chinese LLM ecosystems, including DeepSeek, ChatGLM, and Hunyuan, account for 8 studies (7%). Their presence reflects increasing adoption of regionally developed GenAI technologies and contributes to greater linguistic and technological diversity within immersive educational applications, particularly where multilingual interaction and localized educational content are important.

GAN-based model families remain the least represented, appearing in only 3 studies (3%). Their limited representation is consistent with the increasing prominence of transformer- and diffusion-based generative approaches in recent immersive educational research. Nevertheless, GANs continue to provide value in specialized applications involving synthetic image generation and visual content creation.

Overall, the evidence-derived distribution reveals a heterogeneous landscape of proprietary provider ecosystems, open-source implementations, model families, and region-specific GenAI technologies. Although OpenAI-based technologies remain the most frequently reported, the adoption of alternative ecosystems and generative approaches indicates increasing technological diversification. At the same time, the predominance of proprietary platforms and the substantial number of studies that do not identify the specific LLM used highlight important considerations regarding vendor dependency, reproducibility, operational costs, reporting transparency, institutional control, and equitable access in future AI-driven immersive education systems.

4.4 Degree of freedom (DoF)

The degree of freedom (DoF), extracted from Supplementary Table S7 and summarized in Figure 2e, provides an important indicator of the level of interaction and immersion supported by Metaverse-based educational systems. DoF describes the range of user movement within virtual environments, spanning from limited rotational interaction (3DoF) to unrestricted positional and rotational movement (6DoF). Since immersion and interaction are fundamental characteristics of Metaverse technologies, DoF serves as a useful proxy for evaluating the realism and experiential quality of educational applications.

As illustrated in Figure 2e, 6DoF systems dominate the reviewed literature, appearing in 67 studies (59%). By contrast, 41 studies (36%) do not explicitly report the DoF supported by their platforms. Only a small number of studies employ 3DoF configurations (4 studies, 4%), while 2 studies (2%) fall into the not applicable category. This distribution demonstrates a strong research preference for highly immersive interaction, although reporting practices remain inconsistent across the literature.

The technical contributions summarized in Supplementary Table S7 further reveal that 6DoF systems are predominantly implemented within fully immersive VR, XR, and Metaverse environments that allow users to navigate, explore, and manipulate virtual objects naturally within three-dimensional space. Many of these systems incorporate advanced interaction modalities, including hand tracking, gesture recognition, embodied avatars, and spatial object manipulation, creating rich experiential learning environments. Such capabilities are particularly valuable in simulation-based education, where realistic interaction improves learner engagement, procedural understanding, and practical skill development. Several studies further combine 6DoF immersive environments with LLMs and intelligent virtual agents to provide adaptive tutoring, real-time guidance, collaborative learning, and personalized educational support.

The predominance of 6DoF technologies reflects a broader shift toward embodied and experiential learning, where learners actively interact with virtual environments instead of passively consuming instructional content. This trend is especially evident in healthcare training, engineering education, virtual laboratories, STEM learning, and collaborative educational spaces, where unrestricted movement and realistic interaction substantially enhance immersion, spatial cognition, and knowledge retention.

The large proportion of studies that do not specify their supported DoF represents one of the notable reporting limitations identified in this review. Many of these studies primarily focus on pedagogical frameworks, AI integration, intelligent tutoring systems, or content generation while providing limited technical descriptions of the underlying interaction mechanisms. The absence of standardized reporting makes it difficult to compare immersion levels across systems and reduces the reproducibility of experimental findings. Future research would therefore benefit from consistently documenting interaction characteristics, tracking technologies, and supported movement capabilities.

Although only four studies adopt 3DoF interaction, these systems remain relevant for educational scenarios requiring lightweight deployment, reduced hardware costs, and broader accessibility through mobile devices or web-based platforms. Nevertheless, the absence of positional tracking limits users’ ability to perform natural spatial interaction, making 3DoF environments less appropriate for applications involving complex manipulation, procedural training, or collaborative immersive experiences. Consequently, their use remains comparatively limited within Metaverse-based education.

The observed distribution highlights the continuing evolution of educational Metaverse systems toward highly immersive, interaction-rich learning experiences enabled by 6DoF technologies. At the same time, the substantial proportion of studies with unspecified interaction capabilities underscores the need for more comprehensive reporting standards. Clear documentation of DoF characteristics, interaction modalities, and tracking technologies will facilitate reproducibility, enable meaningful cross-study comparisons, and support the future design and evaluation of scalable immersive learning environments.

4.5 Target users

The distribution of target users, derived from Supplementary Table S8 and illustrated in Figure 2d, provides valuable insight into the primary beneficiaries of Metaverse- and GenAI-enabled educational systems. Understanding the intended user groups helps reveal current research priorities, identify underserved populations, and assess the practical reach of emerging educational technologies.

As shown in Figure 2d, higher education learners represent the dominant target group, appearing in 55 studies (48%). Healthcare and medical trainees constitute the second largest category with 16 studies (14%), followed by K–12 students (13 studies, 11%) and vocational and workforce trainees (11 studies, 10%). Other user groups include educators (9 studies, 8%), language and communication learners (8 studies, 7%), special education learners (6 studies, 5%), industry professionals and domain experts (4 studies, 4%), and other user populations (4 studies, 4%). These findings indicate that current research is primarily oriented toward formal education and professional skills development.

A closer examination of Supplementary Table S8 further reveals that university students across disciplines such as STEM, healthcare, engineering, architecture, social sciences, and language education constitute the primary focus of existing systems. These studies frequently combine immersive VR/XR technologies with LLMs, conversational agents, and adaptive tutoring systems to deliver personalized instruction, collaborative learning, virtual laboratories, and simulation-based educational experiences.

The substantial representation of healthcare and medical trainees suggests growing research attention to immersive technologies for clinical education and professional training. Many studies employ virtual patients, intelligent tutoring systems, AI-assisted diagnosis support, and realistic procedural simulations that enable learners to practice clinical decision-making, communication, and technical skills within safe and controlled virtual environments.

Although K-12 learners and vocational trainees receive comparatively less attention than higher education students, the analyzed publications suggest increasing research attention to these populations. Existing applications primarily support STEM education, language acquisition, foundational science learning, vocational skill development, and technical workforce training through interactive simulations, gamified environments, and personalized AI-generated learning experiences.

Several studies also target educators by providing AI-assisted content generation, lesson planning, assessment support, and classroom management tools. Similarly, language and communication learners benefit from conversational AI systems capable of delivering adaptive feedback, multilingual interaction, pronunciation guidance, and personalized language practice. Emerging research further addresses accessibility through educational systems designed for special education learners, including children with autism spectrum disorder, speech and communication impairments, and other learning disabilities. Nevertheless, these inclusive applications remain relatively limited compared with mainstream educational settings.

The comparatively small number of studies involving industry professionals and other user groups suggests that the adoption of Metaverse and GenAI technologies beyond traditional educational environments is still in its early stages. Expanding future research to include lifelong learners, corporate training, continuing professional education, and diverse learner populations could substantially broaden the societal impact of these technologies.

Overall, the observed distribution demonstrates that Metaverse- and GenAI-enabled educational research remains strongly centered on higher education while progressively expanding toward healthcare, K–12 education, vocational training, and inclusive learning. At the same time, the limited representation of educators, industry professionals, and special education populations highlights important opportunities for future work aimed at improving accessibility, inclusivity, and stakeholder diversity across immersive educational ecosystems.

4.6 First author affiliation country

The geographic distribution of first authors’ affiliations, illustrated in Figure 3, provides important insight into the global research landscape of Metaverse- and GenAI-enabled education. Based on the 114 primary studies included in this review, the research spans 36 countries across North America, Europe, Asia, Africa, South America, Oceania, and the Middle East, demonstrating the growing international interest in immersive and AI-enhanced educational technologies. Nevertheless, the distribution remains highly uneven, with research activity concentrated in a relatively small number of technologically advanced countries.

As shown in Figure 3, the United States is the leading contributor with 21 publications, followed by China with 17 studies and India with 9 studies. Other major contributors include Hong Kong (8), Greece (5), Japan (5), Mexico (5), Canada (4), the United Kingdom (4), Taiwan (3), Morocco (3), Finland (2), Germany (2), Egypt (2), Malaysia (2), and South Korea (2). An additional 20 countries-including Austria, France, Italy, Portugal, Romania, Spain, Switzerland, Ukraine, Pakistan, Qatar, Saudi Arabia, Sri Lanka, Thailand, Turkey, the United Arab Emirates, Ethiopia, Nigeria, Brazil, Ecuador, and New Zealand - each contributed a single publication. These findings demonstrate that while research participation has become increasingly global, the overall output remains dominated by a limited number of countries with mature digital ecosystems.

The strong representation of the United States, China, India, and several European countries reflects sustained national investments in AI, XR technologies, digital education, and research infrastructure. These countries generally benefit from well-established universities, strong industry-academia collaboration, high-performance computing resources, abundant research funding, and broad access to commercial AI platforms and immersive hardware. Such favorable innovation ecosystems facilitate the design, implementation, and empirical evaluation of sophisticated Metaverse-based educational environments and GenAI applications at scale.

European countries collectively constitute a substantial share of the literature despite relatively modest individual publication counts. Their cumulative contribution highlights active interdisciplinary collaboration across Artificial Intelligence, human-computer interaction, educational technology, and immersive learning. Likewise, countries such as Japan, Canada, South Korea, and Hong Kong demonstrate consistent engagement in areas including intelligent tutoring systems, virtual laboratories, collaborative learning, simulation-based education, and human-centered AI applications.

In contrast, Africa, South America, and several parts of the Middle East remain substantially underrepresented. Although countries such as Morocco, Egypt, Qatar, Saudi Arabia, Nigeria, Ethiopia, Brazil, and Ecuador have begun contributing to this emerging field, their research output remains comparatively limited. Existing studies from these regions often focus on context-specific educational challenges, including healthcare training, workforce development, language learning, and culturally relevant educational applications, indicating growing interest despite comparatively lower research capacity.

The observed geographic imbalance also reflects broader structural challenges that extend beyond research priorities alone. Developing regions frequently face infrastructure inequities, including limited access to reliable high-speed internet, cloud computing resources, advanced graphics hardware, XR devices, and high-performance computing facilities required to develop and evaluate immersive educational systems. These limitations are further compounded by the digital divide, where disparities in digital literacy, institutional technological readiness, and access to modern educational technologies restrict both research capacity and large-scale adoption. Furthermore, the substantial financial costs associated with VR headsets, motion-tracking equipment, high-end computing platforms, commercial GenAI services, and cloud-based AI infrastructure create additional barriers for researchers, educational institutions, and learners operating under constrained resources.

These disparities have important implications for the generalizability and inclusiveness of current research findings. Educational systems developed and evaluated primarily within technologically mature regions may not adequately capture the pedagogical needs, infrastructural constraints, linguistic diversity, and socio-cultural contexts of low- and middle-income countries. Consequently, the evidence base may exhibit geographic bias, limiting the transferability of proposed solutions to underrepresented educational environments.

Overall, the geographic distribution demonstrates both the expanding global reach and the uneven development of Metaverse- and GenAI-enabled education research. While the increasing participation of 36 countries reflects growing worldwide interest, research activity remains concentrated in East Asia, North America, and Europe. Addressing this imbalance will require greater international collaboration, equitable research funding, open-source educational platforms, affordable AI and XR technologies, capacity-building initiatives, and stronger partnerships with institutions in developing regions. Such efforts will be essential for ensuring that future Metaverse- and AI-driven educational innovations are accessible, scalable, culturally inclusive, and capable of benefiting learners across diverse socioeconomic and geographic contexts.

5 Limitations of Metaverse and generative AI-based technologies for education (RQ2)

This section synthesizes the key technical limitations reported across the synthesized literature on Metaverse and GenAI-based educational technologies. Based on the aggregation presented in Table 7; Figure 4 and the corresponding definitions summarized in Column 2 of Supplementary Table S9, the identified limitations are organized into six major categories. To strengthen the analytical interpretation of the reviewed evidence, each limitation category is discussed in relation to the dominant technologies (e.g., VR, AR, XR, Digital Twins, LLM-based systems), educational domains, and study characteristics in which these limitations were most frequently observed.

• Empirical design and validation limitations: This category includes limitations related to the design and evaluation of studies and represents the most frequently reported challenge across the reviewed literature. These limitations were predominantly observed in prototype-oriented studies involving immersive VR/XR learning environments and LLM-assisted educational systems evaluated through short-term experiments. Common issues involve small sample sizes, short experimental durations, and reliance on controlled or laboratory-based environments. Several studies lack control groups, baseline comparisons, or standardized evaluation protocols. Additionally, evaluation approaches are often based on subjective self-reported measures, with limited incorporation of objective performance metrics or longitudinal validation. Variability in experimental settings and participant profiles further affects the consistency and reproducibility of reported outcomes. These limitations were particularly common in higher education and STEM-focused studies, where exploratory proof-of-concept evaluations dominated the research methodology. Representative reference(s): Chu et al. (2025), Abdelmagid et al. (2025), Koike and Majima (2025), Ding and Song (2025), Hemminki-Reijonen et al. (2025).
• System maturity and deployment constraints: This category refers to limitations associated with the readiness, scalability, and operational deployment of proposed systems. These issues were frequently reported in studies involving multi-user Metaverse platforms, XR collaborative environments, and institution-specific educational prototypes. Many solutions are implemented as proof-of-concept systems with limited validation in real-world educational settings. Challenges include restricted scalability, lack of robustness under varying operational conditions, and limited integration with institutional infrastructures such as learning management systems. Interoperability and cross-platform compatibility are also frequently underexplored, particularly in web-based Metaverse platforms and collaborative VR environments requiring synchronized multi-user interaction. These deployment-related limitations were especially common in engineering, simulation-based training, and immersive classroom applications. Representative reference(s): Juarez et al. (2025b), Yang et al. (2025), Viju and Terence (2025), Villegas-Ch et al. (2024), Miyawaki et al. (2024).
• AI reliability and data limitations: This category captures issues related to the performance, trustworthiness, and dependability of AI components, particularly in studies utilizing LLMs, conversational agents, adaptive tutoring systems, and multimodal GenAI models. Reported limitations include data scarcity, dataset bias, and limited domain-specific training data. Several studies highlight inconsistencies in AI-generated educational content, semantic interpretation, and automated feedback generation. Hallucination, prompt sensitivity, and unstable response generation were particularly associated with LLM-driven educational assistants and personalized learning systems. In addition, limitations arise from weak model generalization across diverse educational contexts and learner profiles. These challenges were most visible in language learning, scientific tutoring, and AI-assisted instructional design applications, where content accuracy and pedagogical reliability are critical. Representative reference(s): Chu et al. (2025), Nofal et al. (2025), Song et al. (2024), Koike and Majima (2025), Cheng et al. (2024).
• Immersive performance and fidelity constraints: This category includes limitations affecting the quality, responsiveness, and realism of immersive educational experiences. These challenges were primarily observed in VR-, XR-, and Digital Twin-based systems requiring real-time rendering and interaction. Reported issues involve rendering latency, motion tracking inaccuracies, reduced frame rates, and limited sensory realism. Some studies reported reduced visual fidelity and responsiveness on standalone VR headsets, mobile devices, and browser-based Metaverse platforms with constrained computational resources. Interaction design limitations, restricted environmental realism, and network synchronization challenges also affected user immersion and engagement. These issues were particularly evident in simulation-intensive educational applications, including virtual laboratories, collaborative learning environments, and medical or engineering training systems. Representative reference(s): Sirandass et al. (2025), Imam et al. (2025), El-Hajji et al. (2025), Gutiérrez Maquilón et al. (2024), Constantinescu et al. (2024).
• Ethical, bias, and security risks: This category encompasses concerns related to ethics, transparency, fairness, privacy, and system security in AI-driven and Metaverse-based educational environments. These risks were primarily associated with LLM-based educational assistants, adaptive learning systems, and platforms collecting large volumes of user interaction data. Reported issues include bias in generated educational content, lack of explainability in AI decision-making, and privacy concerns associated with learner monitoring and behavioral data collection. Several studies also highlighted concerns regarding fairness, accountability, and culturally sensitive content representation. Security-related limitations included insufficient safeguards for user authentication, data protection, and secure interaction within immersive collaborative environments. These concerns were particularly discussed in studies focusing on personalized learning, intelligent tutoring systems, and socially interactive Metaverse platforms. Representative reference(s): Chu et al. (2025), Nofal et al. (2025), Luo et al. (2024), Vadisetty et al. (2025), Villegas-Ch et al. (2024).
• Limited domain scope and thematic coverage: This category reflects the restricted educational and thematic applicability of many reviewed systems. Several studies focused on highly specialized use cases, such as STEM laboratories, language learning, or immersive visualization tasks, limiting broader educational generalizability. These limitations were particularly common in early-stage exploratory studies and domain-specific Metaverse applications developed for narrow curricular objectives. Broader applications across diverse disciplines, educational levels, and sociocultural contexts remain comparatively underexplored. Variations in curriculum alignment, pedagogical adaptation, and contextual transferability further constrain cross-domain applicability. In addition, K-12 education, humanities, and inclusive learning environments were less frequently represented compared with higher education and technical training domains. Representative reference(s): Lan and Azimi (2025), Zhu et al. (2025), Zhong Y. et al. (2025).

TABLE 7. Distribution of synthesized technical limitation categories across reviewed studies (n = 114).

The detailed study identifiers (S1–S114) for each technical limitation category are provided in Supplementary Table S5. Categories are non-mutually exclusive; individual studies may be assigned to multiple categories. Therefore, cumulative counts and percentages may exceed n = 114 and 100%, respectively.

5.1 Analysis of technical limitations

Table 7 summarizes the distribution of synthesized technical limitation categories across the 114 reviewed studies, providing insights into the principal challenges that currently constrain the development, evaluation, and large-scale deployment of Metaverse- and GenAI-enabled educational systems.

Empirical design and validation limitations constitute the most prevalent category, appearing in 46 studies (40%). These limitations are primarily associated with early-stage VR/XR prototypes, proof-of-concept Metaverse learning environments, and short-term evaluations of GenAI educational systems. Common methodological shortcomings include small sample sizes, limited longitudinal assessment, lack of control groups, insufficient comparative benchmarking, and inadequate validation across diverse educational settings. The prominence of this category demonstrates that a substantial proportion of existing research remains exploratory, with relatively limited evidence regarding long-term learning effectiveness, scalability, and real-world educational impact.

System maturity and deployment constraints are reported in 34 studies (30%). These challenges are particularly evident in collaborative Metaverse platforms, Digital Twin environments, and institution-specific immersive learning systems. Frequently reported issues include limited scalability, interoperability challenges, infrastructure dependency, complex integration with existing learning management systems, and insufficient support for persistent multi-user environments. Such limitations reveal a persistent gap between experimental prototypes and production-ready educational platforms capable of sustained institutional deployment.

AI reliability and data limitations appear in 29 studies (25%), predominantly in systems utilizing LLMs, conversational agents, intelligent tutors, and AI-assisted content generation. The reviewed literature identifies hallucination, dataset bias, prompt sensitivity, inconsistent semantic interpretation, limited explainability, and insufficient domain-specific training data as recurring concerns. These challenges directly affect the trustworthiness and instructional reliability of AI-driven educational systems, particularly in learning contexts where factual accuracy, pedagogical consistency, and learner safety are essential.

Immersive performance and fidelity constraints are identified in 18 studies (16%), primarily within VR- and XR-intensive applications that require high-fidelity rendering and real-time interaction. Reported challenges include hardware performance bottlenecks, network latency, motion-tracking inaccuracies, rendering overhead, reduced frame rates, simulator sickness, and limited sensory realism. These technical constraints continue to influence immersion quality, user comfort, and the overall effectiveness of experiential learning environments, particularly on standalone and resource-constrained devices.

Limited domain scope and thematic coverage are reported in 17 studies (15%). The existing literature remains heavily concentrated on higher education, STEM disciplines, and healthcare training, while comparatively fewer studies investigate humanities, social sciences, arts education, K–12 learning, vocational education, or culturally adaptive educational environments. This narrow thematic focus may limit the generalizability, inclusiveness, and broader educational applicability of current Metaverse and GenAI solutions across diverse learning contexts.

Ethical, bias, and security risks, although less frequently reported, are identified in 9 studies (8%). These concerns are closely associated with AI-driven personalization, learner data collection, automated decision-making, avatar interactions, and persistent virtual learning environments. Commonly discussed risks include algorithmic bias, privacy violations, data misuse, cybersecurity vulnerabilities, consent management, insufficient transparency, and challenges related to accountability in AI-assisted educational systems. As Metaverse platforms become increasingly data-intensive and AI-driven, addressing these ethical and security considerations will become progressively more important for ensuring trustworthy and responsible educational deployment.

Collectively, these findings demonstrate that the reported limitations extend beyond isolated technical challenges and instead reflect the current maturity of the field. While empirical validation and scalable deployment remain the most pressing concerns, reliable AI behavior, immersive system performance, broader educational applicability, and responsible governance also represent critical barriers to widespread adoption. Advancing Metaverse- and GenAI-enabled education will therefore require stronger experimental methodologies, longitudinal evaluations, interoperable and scalable system architectures, robust AI validation procedures, broader domain coverage, and comprehensive ethical and security frameworks that support sustainable real-world educational implementation.

6 User perceptual studies (RQ3)

This section presents a comprehensive analysis of user perceptual studies reported in the reviewed literature on Metaverse- and GenAI–based educational technologies. The identified studies are categorized into ten key perceptual dimensions, reflecting the diverse evaluation criteria employed across immersive and AI-driven learning environments. An aggregated overview of these categories, including their distribution and relative prevalence, is presented in Table 8. Besides, Supplementary Table S10 summarizes the methodological characteristics of the primary studies, including evaluation setups, devices used (e.g., VR headsets, desktop platforms), participant demographics, and study duration.

TABLE 8. Distribution of user perceptual study dimensions (n = 114).

The detailed study identifiers (S1–S114) corresponding to each User Perceptual Study dimension are provided in Supplementary Table S6. Categories are non-mutually exclusive; individual studies may be assigned to multiple categories. Therefore, cumulative counts and percentages may exceed n = 114 and 100%, respectively.

Overall, user-centered evaluation is incorporated in a subset of the reviewed works, indicating that perceptual validation remains secondary to system development and technical implementation. The following subsections discuss each perceptual category in turn.

• Presence and immersion: This category evaluates the extent to which users feel psychologically immersed in virtual learning environments, capturing the sense of “being there,” engagement, and experiential continuity. It reflects how effectively Metaverse systems simulate realistic and interactive educational spaces. Evaluations are typically conducted using presence questionnaires, immersion scales, and qualitative feedback. Representative reference(s): Li et al. (2024), An et al. (2025), Yoo et al. (2025), Zhu et al. (2025), Wang et al. (2026).
• Usability and user experience: This category assesses system ease of use, interface clarity, interaction quality, learnability, and overall user satisfaction. In Metaverse-based education, usability is critical for ensuring accessibility across diverse learner groups, particularly for novice users. Evaluations commonly employ standardized tools such as the System Usability Scale (SUS), task-based assessments, and subjective user feedback. Representative reference(s): Hui et al. (2025), Hemminki-Reijonen et al. (2025), Wong et al. (2024), Li et al. (2024), Lan and Azimi (2025).
• Cognitive load and attention: This category examines the mental effort required to interact with immersive and AI-driven learning systems, as well as how attention is distributed during tasks. High cognitive load may hinder learning efficiency, particularly in complex virtual environments. Evaluation methods include cognitive workload scales (e.g., NASA-TLX), task performance analysis, and attention tracking. Representative reference(s): Chheang et al. (2024), Monjoree et al. (2024), Chen et al. (2025), Xu et al. (2025), Gianni et al. (2025).
• Learning effectiveness and knowledge gain: This category evaluates the educational impact of Metaverse and AI systems, focusing on knowledge acquisition, comprehension, retention, and learning outcomes. Assessments typically involve pre- and post-tests, performance evaluations, and self-reported learning gains, demonstrating the strong emphasis on educational effectiveness in the literature. Representative reference(s): Abdelmagid et al. (2025), Juarez et al. (2025a), Hui et al. (2025), Ding and Song (2025), Yang et al. (2025).
• Motivation and engagement: This category captures users’ interest, enjoyment, and willingness to participate in learning activities within immersive environments. It reflects affective engagement and sustained interaction, which are critical for effective learning experiences. Evaluations often rely on engagement scales, behavioral observations, and self-reported motivation measures. Representative reference(s): Silaloi et al. (2025), Vijaya et al. (2025), Chen et al. (2025), Naderi and Shojaei (2025), Constantinescu et al. (2024).
• Behavioral and performance analytics: This category evaluates user performance and behavioral patterns within the system, including task completion rates, interaction logs, learning progress, and response times. It reflects objective measures of system effectiveness and learner behavior. Representative reference(s): Abdelmagid et al. (2025), Yang et al. (2025), Wong et al. (2024), Yoo et al. (2025), Chheang et al. (2024).
• Technology acceptance: Technology acceptance assesses users’ willingness to adopt and trust Metaverse and AI-based educational systems. It includes perceived usefulness, ease of use, and readiness for integration into real-world learning environments. Representative reference(s): Hui et al. (2025), Ding and Song (2025), Wong et al. (2024), Li et al. (2024), Silaloi et al. (2025).
• AI-specific evaluation: This category focuses on the evaluation of AI-driven components, including GenAI outputs, personalization quality, and system intelligence. It captures aspects such as accuracy, relevance, and reliability of AI-Generated Content. Representative reference(s): Vadisetty et al. (2025), Cheng et al. (2025), Alipour et al. (2025).
• Social and agent perception: This category evaluates users’ perception of virtual agents, avatars, and social interaction within immersive environments. It includes aspects such as co-presence, communication effectiveness, and perceived realism of agents. Representative reference(s): Zhu et al. (2025), Wang et al. (2026), Shah et al. (2025).
• Multisensory perception: This category assesses the integration of multiple sensory modalities (e.g., visual, auditory, haptic) in immersive learning environments. It reflects the extent to which systems provide rich, multimodal experiences. Representative reference(s): Cheng et al. (2024), Gutiérrez Maquilón et al. (2024).

6.1 Analysis of user perceptual studies

The distribution of perceptual dimensions (Table 8) shows that evaluations of Metaverse- and GenAI–enabled educational systems remain largely focused on usability, learning effectiveness, learner engagement, and immersive experience. Usability and user experience (41%) and Learning effectiveness and knowledge gain (39%) are the two most frequently assessed dimensions, emphasizing the importance placed on interaction quality alongside educational outcomes. Most evaluation studies therefore seek to determine whether immersive learning environments are intuitive to use while simultaneously improving learners’ academic performance.

Considerable attention is also devoted to Motivation and engagement (29%) and Presence and immersion (19%), reflecting the importance of realism, interactivity, and sustained learner participation in immersive education. Across the reviewed literature, Metaverse technologies are consistently associated with increased engagement, experiential learning, and learner motivation when compared with conventional digital learning environments. Collectively, these results reinforce the view that the educational value of immersive environments depends not only on content delivery but also on their ability to create psychologically engaging and interactive learning experiences.

To provide a broader perspective on current evaluation practices, the selected studies were further organized into five methodological categories: (i) usability and user experience evaluation, (ii) engagement and motivational assessment, (iii) learning outcome and knowledge gain evaluation, (iv) cognitive and behavioral performance analysis, and (v) affective and social perception assessment. A comparative synthesis of these methodologies, including commonly adopted instruments, participant characteristics, validation strategies, and methodological limitations, is presented in Table 9.

TABLE 9. Comparative synthesis of evaluation methodologies in reviewed studies.

Usability-oriented evaluations dominate the literature, with many studies employing the System Usability Scale (SUS), user satisfaction questionnaires, and task-completion assessments. Learning-oriented evaluations commonly rely on pre- and post-tests, academic performance analysis, and self-reported learning gains. Cognitive evaluations frequently incorporate workload measures such as NASA-TLX, interaction analytics, attention monitoring, and behavioral logging. In contrast, affective and social evaluations-including immersion, co-presence, avatar perception, and emotional engagement-remain comparatively less common despite their importance in immersive educational settings.

More specialized perceptual dimensions receive substantially less attention. Cognitive load and attention (13%) and Behavioral and performance analytics (12%) are only moderately represented, indicating growing but still limited interest in understanding learners’ mental workload, interaction patterns, and task performance. Consequently, current evaluation practices continue to emphasize immediate user perception more strongly than underlying cognitive processes, learning strategies, and long-term behavioral development.

Adoption-related issues also remain relatively underexplored. Technology acceptance accounts for only 7% of the primary studies despite its importance for institutional implementation and sustained use. Factors such as trust, willingness to adopt, long-term usage intention, and organizational readiness are addressed only sporadically. Likewise, AI-specific evaluation appears in just 4% of the literature, even though GenAI and intelligent agents are increasingly integrated into educational Metaverse platforms. Few studies systematically examine the reliability, transparency, explainability, accuracy, or pedagogical appropriateness of AI-Generated Content, leaving an important gap in assessing the educational trustworthiness of these systems.

The least investigated dimensions are Social and agent perception (5%) and Multisensory perception (3%). Collaborative interaction, avatar realism, social presence, and multimodal sensory integration therefore remain insufficiently explored despite being defining characteristics of immersive virtual environments. Existing evaluation practices continue to focus primarily on individual user experience rather than socially collaborative and multisensory learning processes.

Methodologically, most investigations rely on small-scale experiments, short-term deployments, and perception-based assessments. Participants are typically university students or relatively small classroom cohorts, while evaluation periods often range from a single session to only a few weeks. As a result, much of the existing evidence reflects immediate learner perceptions and interaction quality rather than sustained educational effectiveness or long-term pedagogical impact. The widespread dependence on self-reported questionnaires further introduces potential response bias and limits the objectivity of several reported outcomes.

The methodological rigor and external validity of the literature remain uneven. Randomized controlled trials, longitudinal investigations, comparative multi-group studies, and cross-institutional deployments are comparatively rare. Only a limited number of studies benchmark immersive learning environments against conventional instructional approaches using statistically rigorous experimental designs. Standardized evaluation protocols are also applied inconsistently, making direct comparison and replication across studies challenging. Differences in evaluation metrics, participant demographics, technological platforms, and learning scenarios further complicate the synthesis of robust conclusions regarding educational effectiveness.

Another noticeable characteristic of the literature is the distinction between prototype-oriented demonstrations and mature educational deployments. A substantial proportion of studies evaluate newly developed Metaverse platforms, virtual classrooms, AI tutors, or immersive interaction techniques through pilot implementations or laboratory experiments, focusing primarily on feasibility, usability, and initial learner reactions. Far fewer studies investigate mature educational systems operating within authentic classroom settings, curriculum-supported learning activities, or institution-wide deployments. The evidence base therefore remains dominated by proof-of-concept validation rather than large-scale pedagogical implementation and longitudinal educational assessment.

Evaluation maturity also varies across application domains. STEM, healthcare, and safety-training applications generally demonstrate stronger empirical validation because learning outcomes can be measured objectively through task completion, procedural accuracy, error reduction, simulation performance, or pre- and post-test scores. By comparison, disciplines such as the humanities, social sciences, language education, business education, K–12 education, and special education often involve outcomes including communication quality, creativity, collaboration, cultural understanding, accessibility, learner autonomy, and socio-emotional development. These constructs are inherently more difficult to evaluate using short-term usability or engagement measures alone. Accordingly, future evaluation strategies should be aligned with discipline-specific learning objectives rather than relying predominantly on generic usability and engagement metrics.

Despite these methodological limitations, the reviewed studies consistently report improvements in learner engagement, immersion, motivation, and perceived educational value. Nevertheless, the current evidence remains largely exploratory and prototype-driven, with considerably greater emphasis on demonstrating technical feasibility and user acceptance than on rigorously evaluating long-term learning effectiveness, inclusivity, scalability, and knowledge retention. Evaluation maturity therefore remains at an early stage, particularly for GenAI-enhanced Metaverse learning environments.

Taken together, the evidence shows that contemporary user-centered evaluations are dominated by usability, engagement, immersion, and perceived learning effectiveness, whereas AI reliability, technology acceptance, social interaction, multisensory perception, and long-term educational outcomes receive comparatively limited attention. Addressing these gaps will require more comprehensive evaluation frameworks that integrate cognitive, behavioral, social, pedagogical, and AI-specific perspectives.

Such dimensions may also be interpreted through established pedagogical theories, although theoretical alignment should not itself be considered evidence of improved learning outcomes. For example, presence, immersion, and multisensory perception are conceptually consistent with embodied cognition because they capture learners’ spatial, bodily, and perceptual interactions with virtual environments (Makransky and Petersen, 2021; Zhang et al., 2025). Cognitive-load and attention measures can similarly be interpreted through cognitive load theory and may help assess whether immersive fidelity and AI-generated guidance support learning or introduce unnecessary processing demands (Hsu and Lee, 2026). Likewise, social and agent-perception measures relate conceptually to constructivist and social-learning principles through co-presence, peer collaboration, and avatar- or agent-mediated interaction, while behavioral analytics and AI-specific evaluation may support mechanisms associated with self-regulated learning, such as feedback, progress monitoring, and learner reflection (Yan et al., 2025). However, these relationships primarily represent theoretical or design-level alignment unless the corresponding studies empirically demonstrate that the proposed pedagogical mechanisms produce measurable learning gains. Future evaluation frameworks should therefore distinguish theoretical alignment from empirically demonstrated pedagogical effectiveness.

To support the development and validation of such evaluation frameworks, future research would benefit from adopting more rigorous empirical methodologies, including longitudinal studies, controlled comparative experiments, larger and more diverse participant populations, standardized assessment instruments, learning analytics-driven evaluation, and cross-institutional deployments to provide stronger evidence regarding the effectiveness, scalability, inclusivity, and long-term educational value of Metaverse-based learning technologies.

6.2 Learning outcomes: immediate gains, retention, and transfer

Although learning effectiveness and knowledge gain constitute one of the most frequently evaluated dimensions in the reviewed literature (45 studies, 39%; Table 8), the evidence is predominantly based on short-term learning outcomes rather than sustained educational impact. Most empirical studies assess learning gains through immediate pre- and post-intervention comparisons conducted within a single instructional session or over relatively short intervention periods. These evaluation designs provide useful evidence that integrating Metaverse technologies with GenAI can enhance learners’ understanding, engagement, and perceived learning performance immediately following the learning activity. However, such findings primarily demonstrate short-term acquisition of knowledge and should not be interpreted as evidence of durable learning or long-term educational effectiveness.

A more limited number of studies extend evaluation beyond immediate post-tests by employing longitudinal or learning-analytics-based approaches capable of tracking learner progression over time (see Supplementary Table S10). These studies offer preliminary insights into knowledge retention and learning trajectories, yet they remain exceptions rather than common practice within the literature. Consequently, relatively little empirical evidence currently exists regarding whether the knowledge and skills acquired in AI-enabled Metaverse environments are retained after the intervention has concluded.

Similarly, transfer of learning - the ability to apply acquired knowledge or skills to authentic, unfamiliar, or real-world contexts - remains largely underexplored. While several healthcare, vocational, and safety-training applications evaluate procedural performance within realistic simulation environments, most studies stop short of assessing whether these competencies are maintained or successfully transferred beyond the virtual setting. As a result, existing evidence provides only indirect indications of learning transfer, with few studies incorporating authentic workplace assessments, delayed performance evaluations, or longitudinal follow-up capable of demonstrating sustained competence.

Overall, the current body of evidence demonstrates encouraging support for immediate learning gains, learner engagement, and perceived educational value. Nevertheless, robust evidence regarding long-term knowledge retention, transfer of learning, and sustained educational impact remains limited because relatively few studies employ delayed assessments, longitudinal designs, or authentic performance evaluations. Future research should therefore complement traditional usability and satisfaction measures with rigorous educational assessment frameworks that incorporate delayed post-tests, longitudinal learner tracking, competency-based evaluation, and real-world performance measures. Such approaches would provide stronger evidence regarding the enduring pedagogical value of Metaverse- and GenAI-enabled learning environments.

7 Future directions (RQ4)

Based on the future research directions reported across the primary studies (Supplementary Table S9, Column 2), we derive a consolidated taxonomy that captures the dominant research trajectories in Metaverse and GenAI-driven educational systems. The identified directions reflect both technological advancements and pedagogical evolution, addressing key limitations highlighted in Section 5. Figure 5 further illustrates the alignment between these limitations and emerging research priorities. The identified research directions are presented as thematic priorities rather than being classified into short-, medium-, or long-term implementation horizons. Such temporal prioritization would require subjective assumptions that may vary according to technological maturity, institutional readiness, available resources, and regional educational contexts. Accordingly, the proposed taxonomy provides a flexible evidence-based framework that enables researchers and practitioners to prioritize individual directions according to their specific contexts and evolving educational needs.

• Large-scale and longitudinal validation: A major future direction emphasizes the need for robust empirical validation through large-scale deployments and longitudinal studies. Current research is often limited to small sample sizes and short-term usability, engagement, or perception-based evaluations, with limited attention to long-term learning outcomes, knowledge retention, and comparative educational effectiveness. Future work should therefore focus on real-world educational settings, extended study durations, diverse participant populations, and stronger empirical methodologies incorporating control and comparison groups, standardized assessment instruments, and comprehensive learning analytics. These efforts are essential for ensuring generalizability, reliability, and sustained educational impact of Metaverse-based learning systems.
• AI model advancement, trustworthiness, and reliability: Advancing the accuracy, robustness, and trustworthiness of AI models remains a critical priority. Future research should address challenges such as hallucination in GenAI, bias mitigation, explainability, transparency, content verification, and pedagogical reliability. Improving AI trustworthiness is particularly important in educational contexts, where incorrect, misleading, or biased information can negatively influence learning outcomes. Accordingly, future systems should incorporate trustworthy AI validation mechanisms and human oversight strategies to ensure reliable and responsible educational deployment.
• Multimodal and intelligent AI integration: Future systems are expected to integrate multiple AI modalities, including vision, language, speech, gesture recognition, and learning analytics, to enable richer and more natural interactions. This includes the integration of LLMs, computer vision, multimodal sensing, and intelligent tutoring systems to support contextual feedback, adaptive assistance, and immersive educational experiences. Future research should also investigate how multimodal AI can improve learner engagement, collaboration, and real-time educational support within virtual environments.
• Immersive experience, realism, and fidelity enhancement: Enhancing the realism, interactivity, and fidelity of immersive environments is a key research direction. This includes improvements in rendering quality, physics-based interactions, haptic feedback, avatar realism, and sensory immersion. Future work should focus on creating more realistic and engaging virtual learning environments that closely replicate real-world experiences while maintaining usability and performance efficiency across diverse devices and educational contexts.
• Scalability, interoperability, infrastructure, and deployment: Scalability remains a significant challenge, particularly for large-scale educational deployment. Future research should focus on cloud-based architectures, edge computing, efficient rendering pipelines, optimized data streaming, and distributed infrastructures to support scalable and cost-effective Metaverse systems. In addition, interoperability across platforms, devices, and educational ecosystems remains an important research priority. Future systems should enable seamless integration with Learning Management Systems (LMS), digital repositories, and institutional infrastructures to facilitate sustainable deployment and cross-platform educational collaboration.
• Personalization, adaptive learning, and human-in-the-loop systems: Personalized learning is a central goal of AI-driven education. Future work should focus on adaptive systems that tailor content, pacing, interaction, and feedback based on individual learner profiles, preferences, cognitive states, and performance analytics. Moreover, future research should explore human-in-the-loop educational frameworks that maintain educator oversight and intervention within AI-assisted learning environments. Such approaches can support responsible AI usage, improve pedagogical alignment, and ensure that educators remain actively involved in instructional decision-making processes.
• Inclusive, accessible, and culturally adaptive learning environments: Future research should place stronger emphasis on inclusivity, accessibility, and cultural adaptability in immersive educational systems. This includes supporting diverse learner populations, multilingual interaction, accessibility requirements, neurodiverse learners, and resource-constrained educational settings. Future systems should also consider culturally aware content generation and adaptive educational experiences that accommodate varying social, cultural, and technological contexts to promote equitable educational access.
• Domain expansion and interdisciplinary applications: While current studies are concentrated in specific educational domains, future research should explore broader applications across disciplines such as healthcare, engineering, cultural heritage, vocational training, and collaborative professional education. Interdisciplinary integration will enable more diverse, practical, and impactful use cases for Metaverse-based educational systems while facilitating knowledge transfer across multiple application areas.
• Human-centered, pedagogically grounded, and ethically governed systems: Future Metaverse- and GenAI–based educational systems must prioritize human-centered design principles while explicitly grounding immersive learning activities in established educational theories. This includes ensuring usability, accessibility, inclusivity, and alignment with pedagogical objectives through intuitive interfaces, adaptive interaction mechanisms, and support for diverse learner needs. Constructivist and experiential learning principles should guide the design of exploratory, inquiry-based, and simulation-oriented learning tasks, whereas cognitive load theory should inform decisions regarding interface complexity, feedback timing, sensory fidelity, and information sequencing. Embodied cognition principles should further guide the use of gestures, avatars, haptics, spatial interaction, and multisensory experiences so that immersion contributes meaningfully to conceptual understanding rather than functioning merely as technological novelty.

Similarly, social learning and social constructivist perspectives should shape collaborative activities, peer interaction, agent-mediated communication, and community-based knowledge construction, while self-regulated learning should be supported through adaptive scaffolding, metacognitive prompts, learning analytics, and reflective feedback. In parallel, future research must strengthen ethical governance mechanisms addressing privacy, data security, learner wellbeing, algorithmic fairness, transparency, explainability, and responsible AI usage within immersive educational ecosystems. Such pedagogy-driven and ethically grounded system design will be essential for developing trustworthy, sustainable, inclusive, and educationally meaningful immersive learning environments that advance measurable learning outcomes and learner autonomy rather than remaining primarily technology-driven.

Summary perspective: Taken together, the identified future research directions illustrate a clear evolution from prototype-oriented educational systems toward scalable, intelligent, trustworthy, and pedagogically grounded Metaverse learning environments. As illustrated in Figure 5, the synthesized taxonomy demonstrates strong alignment between the principal technical limitations-including empirical design and validation shortcomings, system maturity and deployment constraints, AI reliability and data limitations, immersive performance challenges, limited domain coverage, and ethical, bias, and security risks-and the corresponding priorities for future research.

Several research directions directly target the dominant technical challenges identified across the literature. Large-scale and longitudinal validation addresses the widespread lack of rigorous empirical evidence, while AI model advancement, trustworthiness, and reliability responds to persistent concerns surrounding hallucinations, robustness, transparency, explainability, and instructional accuracy. Complementing these efforts, multimodal and intelligent AI integration aims to enhance adaptive reasoning, richer human-AI interaction, and more capable educational assistants that combine multiple data modalities within immersive learning environments.

The growing emphasis on scalability, interoperability, infrastructure, and deployment further reflects the need to transition from isolated proof-of-concept prototypes to institution-ready educational ecosystems capable of supporting persistent, collaborative, and large-scale learning. At the same time, increasing attention to personalization, adaptive learning, and human-in-the-loop systems recognizes that future Metaverse platforms must provide individualized learning experiences while preserving meaningful educator oversight and pedagogical control.

Beyond technological advancement, the derived taxonomy highlights the importance of inclusive, accessible, and culturally adaptive learning environments together with broader domain expansion and interdisciplinary applications. These directions seek to improve the generalizability of current solutions by extending immersive learning beyond STEM and healthcare into more diverse educational settings while ensuring accessibility for learners with different cultural, linguistic, and educational backgrounds. Finally, the continued emphasis on human-centered, pedagogically grounded, and ethically governed systems reinforces the need to integrate responsible AI principles, learner wellbeing, transparency, privacy, fairness, and sound instructional design throughout the development lifecycle.

Beyond addressing technological challenges, these future research priorities also reinforce the importance of grounding technological innovation in established educational theories. Constructivism and experiential learning continue to motivate the design of authentic, immersive, inquiry-based, and experiential learning environments that actively engage learners in knowledge construction. Social constructivism further emphasizes collaborative virtual environments that facilitate communication, peer interaction, and co-creation of knowledge, while cognitive load theory highlights the need for intelligent interface design, adaptive content presentation, and AI-assisted instructional support that optimize learning efficiency without overwhelming learners. Likewise, self-regulated learning provides an important foundation for personalized learning pathways, intelligent tutoring, adaptive feedback, and learning analytics that empower learners to monitor, regulate, and improve their own learning processes. Collectively, these pedagogical perspectives demonstrate that the long-term success of Metaverse- and GenAI-enabled education will depend not only on advances in immersive technologies and artificial intelligence, but also on their thoughtful integration with evidence-based instructional design principles.

Overall, these future directions indicate a maturing research landscape in which technological innovation is increasingly accompanied by methodological rigor, responsible AI governance, educational inclusiveness, and deployment readiness. Advancing these interconnected priorities will be essential for developing Metaverse learning systems that are reliable, scalable, adaptive, trustworthy, pedagogically grounded, and capable of delivering sustained educational impact across diverse learning contexts through the principled integration of established learning theories and responsible AI.

8 Ethical, governance, and sustainability considerations

Despite the transformative potential of GenAI-enabled Metaverse technologies in education, their adoption introduces ethical, governance, and sustainability challenges that extend beyond those associated with conventional AI-based educational systems. Immersive environments combine GenAI-driven personalization and content generation with persistent virtual spaces, embodied interaction, avatars, spatial computing, and continuous multimodal sensing. Consequently, concerns related to privacy, identity, bias, safety, surveillance, data ownership, accountability, and governance acquire additional dimensions when AI is embedded within immersive and multi-user educational environments. Table 10 summarizes how several conventional AI concerns are extended or amplified within Metaverse and GenAI-enabled education.

TABLE 10. Ethical and governance considerations in conventional AI compared with Metaverse and GenAI-enabled education.

One major concern relates to algorithmic transparency, explainability, and bias (Tabassum et al., 2025; Goktas, 2024). Many GenAI models operate as black-box systems, making it difficult for educators and learners to understand how educational recommendations, automated assessments, generated learning materials, or virtual-agent responses are produced. In immersive environments, these concerns extend beyond conventional model outputs because AI may influence avatar behavior, embodied representation, adaptive virtual interactions, and how learners are represented or treated within virtual spaces. Bias in generated content or AI-driven agents may therefore affect not only information presented to learners but also their social and embodied experiences. Transparent AI decision-making, representative system design, and mechanisms for identifying and correcting biased interactions are consequently important for trustworthy immersive education.

Governance and regulatory frameworks also remain underdeveloped for AI-powered immersive learning ecosystems (Taeihagh, 2025; Calik, 2026). Unlike conventional educational AI applications that may operate within relatively bounded platforms, Metaverse environments can involve persistent multi-user spaces in which learners, educators, avatars, AI agents, platform providers, and generated content continuously interact. This raises additional questions concerning accountability, data ownership, content moderation, acceptable avatar behavior, identity management, platform responsibility, and jurisdiction across shared virtual environments. Governance mechanisms should therefore address not only responsible AI use but also the management of persistent virtual communities, AI-mediated interactions, generated assets, and human–AI participation within educational spaces.

Accessibility, inclusion, and digital inequality represent additional critical concerns (Radanliev et al., 2024; Othman et al., 2024; Yang and Park, 2026). Although Metaverse technologies can create engaging and personalized learning experiences, participation often depends on access to high-performance devices, stable internet connectivity, and adequate digital literacy. These requirements may disproportionately disadvantage learners from under-resourced communities, thereby widening existing educational inequalities. Immersive interfaces may also introduce accessibility barriers associated with headset use, spatial navigation, motion-based interaction, sensory presentation, and embodied control for learners with physical, sensory, or cognitive disabilities. Inclusive design should therefore account for both conventional digital-access barriers and accessibility requirements arising specifically from immersive interaction.

Privacy, surveillance, and learner-data ownership (Golda et al., 2024; Wang et al., 2022; Anidjar et al., 2023; Tukur et al., 2023) are particularly important in immersive educational environments. Unlike conventional educational AI systems that primarily process text, interaction logs, or academic records, Metaverse platforms may continuously collect multimodal and potentially sensitive data, including gaze patterns, body and hand movements, voice interactions, spatial trajectories, biometric signals, behavioral responses, and interactions with virtual objects and agents. When combined with GenAI-based analytics and personalization, continuous sensing may enable detailed inferences about learner behavior, preferences, engagement, and performance, increasing risks associated with surveillance, profiling, and secondary data use. These characteristics raise important questions concerning informed consent, data minimization, retention, ownership, access, and whether learners retain meaningful control over behavioral and biometric data generated through immersive participation. Robust privacy-preserving mechanisms, transparent consent policies, and secure data-governance practices are therefore essential.

Avatar representation and embodied identity introduce additional ethical concerns that are distinctive to immersive learning environments (Lam, 2025). Whereas identity in conventional digital learning systems is typically associated with user accounts or profiles, Metaverse environments allow learners to interact through embodied avatars and potentially persistent virtual identities (Illi and Elhassouny, 2025). Avatar appearance, AI-mediated behavior, identity authentication, impersonation, and representation may therefore influence social interaction, inclusion, trust, and learner experience. GenAI further complicates these issues by enabling automated avatar creation, synthetic voices, virtual agents, and dynamically generated identities or behaviors (Basyoni et al., 2025). Educational platforms consequently require appropriate safeguards for identity integrity, representation, consent, and responsible avatar-mediated interaction.

Another emerging issue involves the ownership and authorship of AI-generated educational content and virtual assets (Mehta and Kaur, 2024; Bozkurt, 2024). GenAI systems can autonomously create learning materials, assessments, avatars, 3D assets, virtual environments, and interactive simulations. In persistent immersive environments, questions of ownership may therefore extend beyond conventional AI-generated text or images to virtual objects and educational assets that can be created, modified, shared, and reused across platforms. Questions remain regarding intellectual property rights, originality, attribution, educational reliability, and ownership of both AI-generated and learner-generated virtual assets. Institutions and policymakers should therefore establish clearer guidelines for content ownership, provenance, verification, attribution, and ethical reuse.

The psychological, cognitive, and immersive safety implications of these environments also warrant careful attention (Makransky and Petersen, 2021; Katernyak and Loboda, 2016). Unlike misinformation encountered through conventional interfaces, misleading, inappropriate, or manipulative AI-generated content may be experienced within highly realistic, spatial, embodied, and socially interactive environments. AI-generated agents and immersive scenarios may therefore influence learners through heightened presence, embodiment, and emotional engagement. Extended exposure may also contribute to cognitive overload, discomfort, reduced physical-world social interaction, or identity-related concerns, particularly among younger learners. Future evaluations should therefore consider not only content accuracy and usability but also psychological safety, cognitive effects, age appropriateness, and the consequences of prolonged immersive exposure.

Finally, sustainability and computational cost considerations remain significant challenges for large-scale deployment (Bashir et al., 2024; Rastogi and Pandita, 2024). Unlike conventional AI applications, GenAI-enabled Metaverse education may combine computationally intensive model training and inference with real-time 3D rendering, multimodal sensing, networking, cloud services, and persistent virtual environments, thereby creating cumulative computational and energy demands. Large-scale adoption may therefore increase infrastructure requirements, operational costs, and environmental impacts if these systems are not designed efficiently. Future research should investigate energy-efficient GenAI models, lightweight immersive architectures, edge and resource-aware computing, and sustainable deployment strategies that balance educational functionality with computational and environmental costs.

Ultimately, responsible GenAI-enabled Metaverse education requires governance approaches that reflect the distinctive characteristics of immersive AI systems rather than relying solely on conventional responsible-AI principles. In particular, privacy and biometric data protection, behavioral surveillance, embodied identity and avatar representation, immersive safety, data and content ownership, accessibility, accountability, multi-user governance, and computational sustainability should be considered throughout system design, deployment, and evaluation. Addressing these interconnected concerns is essential for supporting trustworthy, inclusive, safe, and sustainable immersive educational environments.

8.1 Web3 and blockchain-enabled governance for immersive education

The governance challenges identified above highlight potential opportunities for Web3 and blockchain-based mechanisms to support more trustworthy, transparent, and interoperable GenAI-enabled Metaverse education. In particular, concerns surrounding learner-data ownership, embodied identity, content authorship, provenance, and multi-user governance may benefit from decentralized mechanisms that complement conventional platform-centric approaches. Metaverse environments involve persistent interactions among learners, educators, avatars, AI agents, institutions, and digital assets, creating governance requirements that may extend across organizational and platform boundaries. Web3 infrastructures could therefore provide complementary mechanisms for strengthening trust, accountability, and portability in such environments (Thomason, 2026).

Decentralized identity (DID) and verifiable credentials provide one potential mechanism for strengthening identity management (Peng et al., 2024; Deng et al., 2024). Rather than relying exclusively on platform-controlled accounts, learners could use portable, cryptographically verifiable credentials to establish educational qualifications, institutional affiliations, competencies, or authorized roles while retaining greater control over the information they disclose. Such mechanisms may also support persistent avatar identities by linking virtual representations to verifiable credentials where authentication is required, while selective disclosure can limit unnecessary exposure of personal information (Saiu et al., 2026; Veltri et al., 2026; Thomason, 2026). This is particularly relevant to risks involving impersonation, synthetic identities, and AI-generated avatar behaviors in immersive environments.

Blockchain-based provenance mechanisms may similarly support the traceability and attribution of AI- and learner-generated educational assets (Dai and Wu, 2026; Besiahgari, 2025; Chen et al., 2023). Cryptographic hashes, timestamps, and immutable transaction records could provide auditable evidence concerning the creation, modification, ownership, and transfer of learning materials, avatars, 3D objects, and other virtual assets without necessarily storing the underlying content on-chain (Truong et al., 2023). This could complement institutional policies addressing originality, attribution, intellectual property, verification, and ethical reuse as GenAI-generated assets are increasingly created, modified, and exchanged across persistent environments.

Finally, smart contracts and decentralized governance mechanisms, including decentralized autonomous organizations (DAOs), could support transparent rules for permissions, asset ownership, consent, access, and collective decision-making within shared educational environments (Esposito et al., 2025; Santana and Albareda, 2022). Nevertheless, blockchain should not be regarded as a universal solution; scalability, privacy, interoperability, governance accountability, computational overhead, regulatory compliance, and integration with existing educational infrastructures require careful consideration (Mohammad and Vargas, 2022; Thomason, 2026; Mohammed Abdul, 2024). Future research should therefore empirically examine where Web3 mechanisms provide demonstrable educational and governance value relative to conventional centralized alternatives, particularly in terms of trust, learner autonomy, institutional accountability, and cross-platform interoperability.

9 Discussion and principal findings

The preceding sections presented a structured synthesis of the reviewed evidence across educational application domains, enabling technologies, pedagogical foundations, evaluation practices, technical limitations, governance considerations, and future research directions, as summarized in Figure 6. Building upon these descriptive findings, this section interprets the principal patterns that emerge across the literature. Rather than reiterating the individual taxonomies, the discussion focuses on six cross-cutting tensions that collectively characterize the current maturity of Metaverse- and GenAI-enabled education: technological sophistication versus empirical validation; personalization versus trustworthy AI; immersion versus cognitive load and learner wellbeing; persistent ecosystems versus fragmented prototypes; global aspirations versus geographic and infrastructural concentration; and human-centered design versus limited accessibility and inclusivity evidence. These tensions highlight the broader educational, technological, and methodological implications of the reviewed evidence and the conditions required for progression from experimental systems toward mature educational ecosystems.

9.1 Technological sophistication versus empirical validation

The most prominent tension emerging from the reviewed literature is the contrast between rapidly advancing technological capabilities and comparatively limited empirical evidence of sustained educational effectiveness. The 114 included studies demonstrate increasing integration of immersive technologies and GenAI for personalized tutoring, intelligent content generation, adaptive feedback, conversational agents, collaborative learning, simulation-based training, and experiential education. VR remains the predominant immersive modality, while OpenAI-based models constitute the most frequently adopted GenAI ecosystem, illustrating the rapid convergence of immersive visualization and increasingly capable generative models. However, this technological sophistication is not yet matched by equivalent maturity in educational validation.

Empirical design and validation limitations constitute the most frequently reported category of technical and methodological limitation. Many systems remain prototype-oriented and are evaluated through relatively small-scale or short-term studies, while evidence concerning longitudinal learning effectiveness, knowledge retention, transfer of learning, real-world effectiveness, and institution-wide deployment remains comparatively limited. Although 68% of the reviewed studies reported user perceptual evaluations, these evaluations predominantly emphasize short-term dimensions such as usability, engagement, presence, immersion, and technology acceptance rather than sustained educational outcomes. Consequently, positive user experience should not necessarily be interpreted as evidence of improved learning effectiveness. This distinction is particularly important because technologically sophisticated immersive systems may produce strong perceptions of novelty, presence, or engagement without necessarily demonstrating corresponding improvements in knowledge acquisition or retention.

This imbalance represents a central maturity gap in the field: technological capabilities are advancing faster than the educational evidence required to establish their sustained pedagogical value and real-world readiness. Addressing this gap requires a shift from predominantly proof-of-concept demonstrations toward longitudinal, comparative, and ecologically valid evaluations grounded in clearly articulated pedagogical objectives and established learning theories. Future studies should therefore evaluate not only whether Metaverse and GenAI systems function effectively or are positively perceived, but also whether they produce measurable and sustained educational benefits relative to appropriate instructional alternatives.

9.2 Personalization versus trustworthy AI

A second tension concerns the growing emphasis on AI-driven personalization and adaptive learning alongside unresolved challenges in AI reliability, transparency, bias, privacy, and human oversight. Across the reviewed studies, GenAI increasingly supports personalized tutoring, adaptive feedback, automated educational content generation, conversational interaction, intelligent virtual agents, and context-sensitive learning experiences. These capabilities create opportunities to tailor educational content, instructional pacing, feedback, and learning pathways to individual learners and therefore represent an important direction for future intelligent educational environments.

However, increasing personalization also increases dependence on the reliability of AI-generated outputs and the quality of the learner data used to produce them. The reviewed literature identifies AI reliability and data limitations among the major technical challenges, including hallucinations, inaccurate or inconsistent outputs, insufficient domain knowledge, bias, limited explainability, and dependence on training and interaction data. These concerns become particularly consequential in educational settings because incorrect or misleading generated content may directly influence learner understanding, assessment, and decision-making. Greater personalization may also require increasingly detailed learner modeling and behavioral data, intensifying concerns regarding privacy, profiling, consent, and appropriate data governance.

The pedagogical benefits of personalization must therefore be balanced against requirements for trustworthy AI and meaningful human oversight. Rather than treating personalization as an inherently beneficial outcome of increasingly autonomous GenAI systems, future educational architectures should incorporate mechanisms for content verification, uncertainty awareness, explainability, bias mitigation, privacy protection, and teacher or learner intervention. Human-in-the-loop approaches are particularly important for maintaining pedagogical accountability and ensuring that adaptive systems augment rather than displace educator judgment. Thus, the progression toward personalized Metaverse education should be understood not simply as increasing AI autonomy, but as developing adaptive systems in which personalization, reliability, transparency, and human agency are jointly optimized.

9.3 Immersion versus cognitive load and learner wellbeing

A third tension arises between the educational opportunities associated with increasingly immersive environments and their potential cognitive, psychological, and usability costs. VR, AR, MR, XR, and related immersive technologies can support experiential learning, embodied interaction, spatial understanding, simulation, collaboration, and learning-by-doing. These affordances align particularly well with constructivist, experiential, embodied, and social learning perspectives, where learners actively interact with content and environments rather than receiving information passively.

Nevertheless, greater immersion does not necessarily imply greater educational effectiveness. The reviewed evidence identifies immersive performance and fidelity constraints, including latency, rendering quality, interaction limitations, hardware constraints, and user discomfort, while the broader literature also raises concerns regarding cognitive overload, reduced physical-world interaction, digital dependency, and the psychological effects of prolonged immersive exposure. From a cognitive load perspective, highly interactive and visually complex environments may introduce substantial extraneous processing demands when immersive features are not directly aligned with instructional objectives. Similarly, increasingly realistic or emotionally intense virtual experiences may affect learners in ways that conventional usability and engagement measures do not adequately capture.

The relevant design objective is therefore not maximum immersion, but pedagogically appropriate immersion. Immersive features should be selected according to their contribution to learning objectives, learner characteristics, accessibility requirements, and cognitive demands. Future evaluations should consequently move beyond presence and engagement alone to examine cognitive load, knowledge retention, learner wellbeing, sustained exposure, and the relationship between degrees of immersion and educational outcomes. Such evidence would help distinguish situations in which immersion provides genuine pedagogical value from those in which technological complexity primarily increases cognitive or operational burden.

9.4 Persistent educational ecosystems versus fragmented prototypes

A fourth tension concerns the long-term vision of persistent, interoperable, and intelligent educational ecosystems compared with the fragmented and prototype-oriented nature of much of the current evidence. The reviewed literature increasingly envisions educational Metaverse environments in which learners, educators, intelligent agents, virtual campuses, Digital Twins, immersive simulations, and adaptive GenAI services operate across persistent and interconnected digital spaces. Such environments could support continuity of learning, personalized learner models, collaborative experiences, real-time adaptation, and integration between physical and virtual educational resources.

In practice, however, system maturity and deployment constraints remain among the most frequently reported limitations. Many existing systems are implemented as isolated prototypes designed for specific courses, tasks, technologies, or experimental settings, with comparatively limited evidence concerning interoperability, scalability, cross-platform integration, persistent learner data, institutional integration, and long-term maintenance. This creates a substantial gap between the ecosystem-level vision of the educational Metaverse and the implementation maturity demonstrated by current systems.

Economic and infrastructural feasibility further reinforce this tension. Large-scale deployment may require substantial investment in XR hardware, high-performance computing, cloud services, GPU-intensive GenAI models, high-bandwidth networking, software licensing, cybersecurity, technical support, and ongoing system maintenance. Persistent immersive environments and GenAI services may also generate considerable computational and energy demands, introducing sustainability concerns alongside technical scalability. Future research should therefore treat scalability as a multidimensional challenge involving not only computational performance, but also interoperability, institutional integration, economic feasibility, maintainability, cybersecurity, and environmental sustainability. Progress toward persistent educational ecosystems will depend on modular and interoperable architectures, lightweight and resource-aware technologies, standardized integration mechanisms, and validation under realistic institutional conditions.

9.5 Global educational aspirations versus geographic and infrastructural concentration

A fifth tension concerns the global ambitions associated with Metaverse and GenAI-enabled education relative to the geographic and infrastructural concentration of the current evidence. The reviewed studies span multiple countries and educational contexts, demonstrating growing international interest in immersive and AI-enabled learning. Nevertheless, the distribution of research remains geographically uneven, with a substantial proportion of studies originating from a comparatively limited group of technologically and research-intensive countries. Application domains are similarly concentrated, with STEM, higher education, and healthcare representing prominent areas of investigation.

This concentration has implications for the generalizability of existing findings. Educational practices, technological infrastructure, institutional capacity, language, cultural context, digital literacy, and access to high-performance computing and XR hardware vary considerably across regions. Systems developed and evaluated in technologically well-resourced institutions may therefore not transfer directly to settings characterized by limited connectivity, restricted access to immersive hardware, lower computational capacity, or different pedagogical and cultural requirements. The economic and infrastructure demands associated with XR hardware, cloud services, networking, and computationally intensive GenAI may further widen existing digital inequalities if accessibility and resource constraints are not explicitly considered.

Consequently, the global potential of Metaverse and GenAI-enabled education should be evaluated through more geographically diverse and context-sensitive research. Future studies should expand beyond currently dominant regions and institutional settings and investigate lightweight, affordable, multilingual, and resource-aware solutions appropriate for diverse educational contexts. Cross-cultural and multi-institutional evaluations would also strengthen the external validity of reported findings and clarify whether observed benefits generalize across learner populations, educational systems, and socioeconomic conditions. Global scalability should therefore be understood not only as the technical ability to support larger numbers of users, but also as the capacity to deliver educational value equitably across heterogeneous technological and social environments.

9.6 Human-centered design versus limited accessibility and inclusivity evidence

Finally, the literature increasingly emphasizes learner-centered, adaptive, engaging, and human-centered educational experiences, yet empirical evidence concerning accessibility, inclusivity, and diverse learner needs remains comparatively limited. Many reviewed systems evaluate user experience through measures such as usability, engagement, presence, immersion, satisfaction, motivation, and technology acceptance. These dimensions are valuable for understanding how learners perceive and interact with emerging systems and demonstrate that user-centered evaluation has become an important component of Metaverse and GenAI research.

However, human-centered design extends beyond positive usability or engagement. A genuinely inclusive educational ecosystem must also account for differences in physical ability, sensory requirements, cognitive characteristics, language, digital literacy, socioeconomic resources, age, cultural background, and access to appropriate hardware and connectivity. The reviewed evidence provides comparatively limited systematic evaluation of these dimensions, particularly across diverse learner populations and realistic long-term educational settings. This limitation is especially important for immersive technologies, where hardware requirements, interaction modalities, motion-related discomfort, interface complexity, and physical accessibility can create additional barriers to participation.

Accordingly, accessibility and inclusivity should be treated as core design and evaluation requirements rather than secondary deployment considerations. Future research should involve diverse learner populations throughout system design and evaluation, apply participatory and human-centered approaches, and assess whether adaptive AI and immersive interfaces genuinely reduce rather than reproduce barriers to learning. Evaluation frameworks should therefore extend beyond usability and engagement to include accessibility, equitable participation, learner agency, trust, safety, sustained educational outcomes, and the distribution of benefits across different user groups.

Taken together, these six tensions indicate that the next stage of Metaverse and GenAI-enabled education will depend less on demonstrating additional technological capability and more on establishing when, how, and under what conditions these capabilities generate trustworthy, inclusive, scalable, and sustained educational value. The field is clearly progressing toward more intelligent, adaptive, immersive, and interconnected learning environments; however, technological advancement alone is insufficient to establish educational maturity. Future progress requires stronger alignment among technological innovation, pedagogical theory, empirical validation, human-centered design, responsible AI governance, accessibility, economic feasibility, and sustainable deployment. The future research directions synthesized in Section 7 should therefore be interpreted as interconnected priorities for closing these maturity gaps rather than as independent technological development pathways.

9.7 Strengths and limitations

One of the principal strengths of this study lies in its systematic, transparent, and reproducible review methodology, guided by the PRISMA Extension for Scoping Reviews (PRISMA-ScR) guidelines (Tricco et al., 2018). The review followed a structured workflow encompassing study identification, screening, eligibility assessment, data extraction, and thematic synthesis. To promote transparency and reproducibility, the completed PRISMA-ScR checklist, detailed search strategies, and supplementary replication materials are provided alongside the manuscript. Furthermore, the expanded search strategy, including abstract-level searches in databases such as IEEE Xplore and the ACM Digital Library where appropriate, substantially improved the comprehensiveness of the evidence base and reduced the likelihood of omitting relevant studies. The public availability of the replication materials also facilitates independent verification and future updates of the review, while the inclusion of studies spanning diverse educational domains, technologies, learner populations, and geographical regions enhances the broader applicability of the synthesized findings, notwithstanding the limitations discussed below.

Another important strength is the breadth and depth of the analytical synthesis. Rather than merely cataloguing published work, the review integrates multiple complementary perspectives, including application domains, Metaverse technologies, GenAI categories, pedagogical foundations, target users, technical limitations, user evaluation practices, ethical and governance considerations, and future research directions. The explicit synthesis linking identified technical and methodological challenges with corresponding future research priorities provides practical guidance for researchers, educators, developers, and policymakers seeking to design more scalable, trustworthy, and learner-centered Metaverse education systems.

A further contribution of this review is its comprehensive analysis of evaluation practices across the literature. Beyond summarizing commonly adopted usability and user experience measures, the review identifies important gaps in empirical validation, including the limited use of longitudinal studies, objective learning outcome assessments, cognitive and psychological evaluations, inclusivity measures, and ethical impact assessments. By highlighting these methodological shortcomings, the review provides an evidence-based foundation for improving future evaluation frameworks for AI-driven Metaverse learning environments.

In addition, this review systematically characterizes the maturity of the available evidence. Although no formal methodological quality or risk-of-bias assessment was undertaken, consistent with the objectives of a scoping review, evidence maturity was evaluated through multiple complementary dimensions, including technology maturity (envisioned versus implemented systems), implementation maturity (prototype versus validated systems), evaluation methodologies, validation approaches, DoF, and user study characteristics. These indicators provide readers with important context for interpreting the overall maturity and robustness of the current evidence base while remaining consistent with PRISMA-ScR guidance. However, because methodological quality and risk of bias were not formally assessed, the frequencies and distributions of the reported findings should not be interpreted as evidence of comparative effectiveness, methodological superiority, or strength of evidence.

The reported frequencies and percentages are intended as descriptive summaries of the 114 included studies rather than inferential estimates of a broader population. Several analytical dimensions, including Metaverse technologies, GenAI categories, application domains, target user groups, technical limitations, and user evaluation categories, employed non-mutually exclusive multi-label coding, whereby an individual study could be assigned to multiple applicable categories. Consequently, cumulative counts and percentages across these dimensions may exceed n = 114 and 100%, respectively, and should not be interpreted as mutually exclusive distributions. Percentages for individual categories were calculated independently using the 114 included studies as the denominator. Accordingly, the reported frequencies should be interpreted as indicators of the relative distribution of current research trends and evidence characteristics rather than precise estimates of prevalence or statistical significance. Consistent with the objectives of a PRISMA-ScR, no inferential statistical analyses (e.g., confidence intervals) were performed.

Despite these strengths, several limitations should be acknowledged. First, although the search strategy was designed to maximize coverage across five major academic databases (ACM Digital Library, IEEE Xplore, ScienceDirect, SpringerLink, and Scopus), the review was restricted to English-language, peer-reviewed publications and excluded grey literature, including technical reports, industrial white papers, dissertations, preprints, and other non-peer-reviewed sources. Consequently, relevant studies published in other languages, disseminated through non-peer-reviewed venues, or for which full texts could not be accessed may not have been captured, and some emerging industrial developments, prototype systems, or rapidly evolving applications may therefore be underrepresented. This restriction may also have introduced language and publication bias by disproportionately representing evidence from English-language publications, potentially limiting the global representativeness of the synthesized findings, particularly for regions where high-quality research is published in other languages.

Second, database-specific search capabilities and indexing mechanisms required slight adaptations of the search strategy across different digital libraries. Although the protocol was carefully designed and expanded to include abstract-level retrieval where feasible, variations in database indexing, metadata quality, controlled vocabularies, and search functionalities may still have resulted in some relevant studies being overlooked, particularly those employing unconventional terminology, newly emerging concepts, or publications not indexed within the selected databases.

Third, the dominance of conference publications among the included studies reflects the rapidly evolving nature of this research area. While conference papers provide timely insights into emerging technologies and early-stage innovations, many describe prototype systems or short-term evaluations, which often provide less comprehensive empirical evidence than mature journal publications. Consequently, the current evidence base remains weighted toward proof-of-concept implementations rather than large-scale, longitudinal educational deployments.

Although study selection, data extraction, and thematic coding were independently conducted and cross-validated by multiple reviewers, followed by discussion and consensus to resolve disagreements, some degree of subjective interpretation is unavoidable. In particular, the taxonomies developed through open coding and thematic synthesis represent author-derived analytical interpretations of heterogeneous studies spanning diverse AI technologies, XR platforms, pedagogical approaches, and evaluation methodologies. Although operational definitions, coding rules, independent review, and consensus procedures were employed to improve consistency, alternative classification schemes may therefore also be possible.

To further evaluate the consistency of the study selection process, a post hoc inter-rater reliability assessment was conducted in which an independent reviewer re-screened a randomly selected sample of 110 full-text articles (approximately 30% of the 368 articles assessed for eligibility) using the original inclusion and exclusion criteria. Agreement between the original and independent screening decisions yielded a Cohen’s κ value of 0.834, indicating almost perfect agreement according to the Landis and Koch classification. Although the high level of agreement supports the consistency of the screening decisions, the retrospective reliability assessment does not substitute for a prospectively planned inter-rater reliability assessment and should therefore be considered a methodological limitation. To further enhance transparency and reproducibility, the list of the 110 randomly selected articles used in the post hoc reliability assessment is provided in the Replication File included in the Supplementary Material.

A further limitation concerns the GenAI classification, which partly reflects how technologies are reported in the primary studies. Consequently, the resulting categories may represent provider or model ecosystems, open-source platforms, model families, or unspecified reporting rather than a uniform architectural classification. This reporting-oriented taxonomy was retained because the review aims to characterize the GenAI technologies and ecosystems explicitly reported or implemented across the included studies; nevertheless, differences among these categories should not be interpreted as representing directly equivalent technological distinctions.

Finally, the review intentionally focuses on studies published or indexed from January 2022 up to the respective database search dates in 2026 to capture the rapid emergence of GenAI within Metaverse-enabled education. While this temporal scope provides a current snapshot of the field, the exceptionally rapid evolution of GenAI models, platforms, and immersive technologies means that some technological classifications, adoption patterns, and implementation trends may become outdated as new systems emerge. The findings should therefore be interpreted as a time-bounded representation of the evidence available up to the respective search dates, and periodic updates will be necessary as the field evolves. Furthermore, the selected temporal scope necessarily excludes earlier foundational research on immersive learning, virtual environments, and intelligent tutoring systems that may provide additional historical context regarding the evolution of contemporary educational Metaverse technologies.

Despite these limitations, the review provides a comprehensive, rigorous, and up-to-date synthesis of the rapidly evolving intersection between GenAI and Metaverse technologies in education. By integrating technological, pedagogical, methodological, evaluative, ethical, and governance perspectives within a unified analytical framework, the study establishes a robust foundation for future research and provides practical guidance for the development of more scalable, trustworthy, inclusive, and sustainable AI-enabled Metaverse learning ecosystems.

10 Conclusion and future outlook

This scoping review synthesized evidence from 114 primary studies on Metaverse and GenAI-enabled education published or indexed from January 2022 up to the respective database search dates in 2026. The findings show increasing integration of immersive technologies and GenAI across diverse educational contexts, particularly STEM, higher education, and healthcare. VR was the predominant immersive modality (61%), while OpenAI-based models represented the most frequently reported GenAI ecosystem (60%). Across these studies, GenAI was commonly integrated with immersive environments to support functions such as intelligent tutoring, adaptive feedback, virtual agents, personalized learning, and automated content generation.

The principal contribution of this review is an integrative, multidimensional synthesis of the field across application domains, Metaverse technologies, GenAI categories, pedagogical foundations, target users, evaluation practices, technical limitations, and future research directions. This synthesis reveals a notable gap between technological advancement and empirical maturity. Many reported systems remain prototypes or proof-of-concept implementations, while recurring limitations include insufficient empirical validation, system maturity and deployment constraints, AI reliability and data limitations, immersive performance constraints, scalability and interoperability challenges, accessibility barriers, and ethical and governance concerns.

User evaluation represents a particularly important evidence gap. Although approximately 68% of the reviewed studies reported user perceptual evaluations, these assessments predominantly focused on short-term dimensions such as usability, engagement, satisfaction, presence, and immediate learning experiences. Comparatively fewer studies examined longitudinal learning effectiveness, knowledge retention, accessibility, inclusivity, AI trustworthiness, or sustained real-world deployment. Taken together, these findings indicate that technological sophistication in Metaverse and GenAI-enabled education is currently advancing faster than the educational evidence needed to establish sustained learning effectiveness, inclusivity, trustworthiness, and real-world readiness.

Several issues therefore remain unresolved. More robust evidence is needed from large-scale and longitudinal educational deployments, together with standardized evaluation approaches capable of assessing sustained learning outcomes across diverse learners and authentic educational settings. AI reliability, transparency, human oversight, accessibility, interoperability, privacy, governance, and institutional deployment also remain important challenges. Addressing these issues will be necessary to determine whether the technological capabilities demonstrated by current systems translate into sustained and broadly applicable educational benefits.

Overall, this review provides an evidence-based framework for understanding the current state and maturity of Metaverse and GenAI-enabled education. By integrating technological, pedagogical, methodological, evaluative, ethical, and governance perspectives, it identifies both the capabilities demonstrated by the existing evidence base and the areas where stronger empirical support remains necessary. Future progress should therefore prioritize rigorous educational evidence alongside technological development to support trustworthy, inclusive, and scalable adoption in real-world educational settings.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. The author(s) declare that Generative AI was used in the creation of this manuscript. During the preparation of this work, the author(s) used ChatGPT in order to improve readability and language. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fbloc.2026.1860582/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Jane Thomason, University College London, United Kingdom

Reviewed by: Lidia Bajenaru, National Institute for Research and Development in Informatics, Romania

Monica Barbu, National Institute for Research and Development in Informatics, Romania

Footnotes

Note: The “Applied Filters” column denotes database-specific constraints (e.g., publication period, document type, language, and search field). The initial searches covered the period from January 2022 to February 2026, while the expanded abstract-level searches of IEEE Xplore and ACM Digital Library were conducted in June 2026. Full database-specific search dates, fields, filters, query strings, and retrieval counts are provided in the Supplementary Material (Digital_Libr_Info_Full).

The study addresses the integration of Metaverse/XR and GenAI technologies within an educational context, with both technological components being clearly identifiable and meaningfully incorporated into the investigated system, application, framework, or educational experience. Studies were not required to explicitly use the term “GenAI,” provided that the employed AI technology could be clearly identified as generative in nature.