When systematic reviews accumulated, surrounding the application of AI technologies in educational domains, researchers proposed comprehensive frameworks for reviewing emerging technologies. From the perspective of research practice, such frameworks pointed out critical aspects that review articles should aim to cover partially or fully. As an early comprehensive review, Roll and Wylie’s (2016) article included research foci, disciplines, learner-computer interaction patterns, technological settings, learning objectives, and the expected changes in education. A more recent review framework suggested literature synthesis regarding the applications, samples, research methods, roles of AI, algorithms, and research topics (Hwang and Tu, 2021; Liang et al., 2023). From the perspective of research topic conceptualization, such review frameworks had assisted in establishing the contributions of review articles that narrowed to specific contexts, such as technologies, disciplines, and educational levels (see Huang et al., 2023 for a recent review on AI in language education). Zawacki-Richter et al.’s influential systematic review (2019) refined four primary roles of AI in education, raising discussion on the transformation of instructors’ roles.

Educational quality and assessment were critical in traditional and technology-enhanced educational modes. Researchers adopted surveys to measure perceptual and attitudinal outcomes, such as motivation, engagement, acceptance, perceived usefulness, and intentions to use through self-reported perceptual data (Lei et al., 2024; Sallam et al., 2023). In contrast, many studies implemented intervention designs and standardized tests to evaluate the improvements in knowledge acquisition, retention, and skill learning when learners were assisted by particular educational technologies (Lee and Jeon, 2024). In the vast body of literature on pedagogical designs and educational technologies, multifaceted effectiveness measurements were adopted to reflect whether certain innovative teaching modes were better than the traditional class without the investigated designs (Lin and Yu, 2023; Wu and Yu, 2024). The indicators reflected Bloom’s taxonomy of learning objectives in knowledge, skills, and affective aspects (Bloom, 1956; Hoque, 2016) but delved into finer dimensions.

The embodiment principle in technology-enhanced education suggested that people could learn better with human-like elements in learning media, including gestures, movements, and voices (Clark and Mayer, 2016). The effects of embodiment on learning effectiveness could be associated with social presence (Lin and Yu, 2025; Zhang et al., 2022). Social presence refers to the feeling of personal presence in specific environments, which is a critical concept in interactive educational technologies (Oh et al., 2018). Embodied elements in digital learning environments could provide learners with interaction and presence in the real world with human beings. Numerous studies explored pedagogical agents in education, such as the effects of facial expressions, emotional feedback, and communication with students (Schneider et al., 2022; Sikström et al., 2022).

In the theoretical discussion on technology integration into educational domains, frameworks were proposed to capture the elements of technology-enhanced education and their interaction. At the macroscopic level, researchers examined the extent to which technology integration transformed the learning modes. The Analysis, Design, Development, Implementation, and Evaluation (ADDIE) model guided many researchers to evaluate and improve technology-enhanced education practices (Luo et al., 2024). The Replacement, Amplification, and Transformation (RAT) model was proposed for teachers to examine their abilities to incorporate educational technologies into their teaching practice (Hughes et al., 2006). It demonstrated the intended purposes of technology applications in pedagogical practice from the perspective of teachers’ roles.

At the microscopic level, technology-enhanced education could be interpreted as the interaction between elements, such as students, teachers, learning contents, and technologies. In pedagogical theories that did not involve modern educational technologies, constructionists considered learning to be the interaction between learners, teachers, and learning contents (Anderson, 2003; Piaget, 1970). Examining educational outcomes and standards of e-learning, Barari et al. (2022) adopted the interaction model between teaching methods, learning theories, and technologies, which was adapted from Dabbagh’s prototype (2005). In response to the review frameworks of AI in education, this study further specified research methods, application contexts, design features, effectiveness, and roles of AI-generated embodied virtual teachers in our integrated working conceptual framework (Fig. 2).

An emerging category known as Generative AI was built on Large Language Models (LLMs) and utilized deep learning technologies for training with an enormous size of data (Nah et al., 2023). It generated textual content, music, images, and videos creatively based on diverse sources (Chen et al., 2024; Dasborough, 2023). Much earlier than the new era when AI could generate complete video clips, researchers and instructors benefited from the intelligence synthetic techniques in incorporating generated embodied virtual teachers into instructional videos (e.g., Bergmann and Macedonia, 2013). AI-generated learning partners engaged learners in virtual calls to practice their English-speaking skills, which was among the popular and emerging mobile applications (Wan and Moorhouse, 2024). With a similar technical basis, the generated virtual teachers took the simulated, dynamic, and vivid appearances close to human teachers, and they could enhance learning outcomes as the embodied learning theories suggested (Macedonia, 2019). Early technological practitioners only controlled the generation and synthesis of large chunks of virtual teachers, while more recent technologies allowed finer adjustments, such as automated lip synchronization (Alshahrani and Maashi, 2024). Empirical evidence showed that generated embodied virtual teachers in learning videos had inconsistent impacts on learning. Some suggested that intrinsic motivations were enhanced, while others reported the nature of the short-term novelty effect (Deng et al., 2022; Vallis et al., 2024). With the advancements in generative AI technologies, it was imperative for researchers to clarify the effects of AI-generated embodied virtual teachers and seek solutions to the challenges that might confound future applications.

Dominant existing reviews examined AI chatbots and robots, with a dearth of research calling for closer examinations of AI-generated embodied virtual teacher presence. Wu and Yu’s review (2023) provided inspiring evidence for AI chatbots in education, including their roles as teaching assistants and instructors, but they did not particularly concentrate on the roles of pedagogical agents. Armando et al.’s review (2022) examined a particular aspect of pedagogical agents but did not concentrate on AI-empowered agents despite the enlightening implications for pedagogical practice and technological designs. Davis et al. (2021) suggested that a high “persona rating” (p.89), or perceived embodiment of the virtual teachers, did not indicate more positive impacts on learning outcomes. However, they identified facial expressions and gestures as significant embodiment features, shedding light on embodied virtual teacher designs (Davis et al., 2021). Meta-analytical evidence supported that embodied pedagogical agents significantly enhanced learning outcomes when testing formats and learner strategies were considered (Davis et al., 2023). Intimacy and human likeness might be achieved through embodied designs, while AI allowed natural language comprehension and emotional comprehension even in conversational agents without digital human appearances (Ortega-Ochoa et al., 2024). As Table 1 summarizes, few reviews concentrated on virtual pedagogical and humanoid agents generated by AI. The rise of AI in education necessitated a comprehensive literature synthesis to understand this topic.

The advent of AI technologies restructured the roles of educational stakeholders, including students, teachers, institutional administrators, advisors, and researchers. In traditional classrooms, teachers were expected to draft teaching objectives, prepare materials, deliver lectures, and provide assignment feedback. However, a comprehensive review of AI in higher education indicated that AI undertook learner profiling, automated evaluation, and personalization (Zawacki-Richter et al., 2019). Similar to the emergence of computers in classrooms, teachers’ roles transformed from imparters to coaches and facilitators of students’ technology use (Hannafin and Savenye, 1993). The increasingly crucial roles of AI urged researchers to explore the trajectories whereby human instructors and teachers might collaborate with technologies in the AI era. Li et al. (2022) established that the preferable modes of using AI in education were interdependent and exploratory. The former suggested that human beings should actively dedicate themselves to tasks that AI could not perform, and the achievements of humans and AI should rely on each other to yield better results. The latter suggested adopting AI technologies to create and organize educational resources better.

Finer aspects of the human-AI collaboration modes implied redistributed roles of teachers and students in educational contexts. Based on 92 articles, Chiu et al. (2023) confirmed AI’s advantages perceived by teachers, encompassing reduced workload and enhanced teaching competence. However, their findings also indicated teachers’ confusion about current collaboration modes with AI in education, concluding with teachers’ concerns about ethical aspects and teaching effectiveness. Particularly, teachers expressed concerns about pedagogical agents, considering them insufficiently compatible with diverse teaching methods, designs, and contexts and complaining about their poor performances in explaining reasons to students (e.g., Holstein et al., 2019). Despite the unwillingness to acknowledge the trend of technology-enhanced education, especially in higher education, AI-generated embodied virtual teachers might replace human instructors, at least partially in delivering lectures. It was imperative to synthesize the existing literature on this technology to clarify how well it performed and how human instructors should react.

Grounded on the review framework and the underexplored issues related to AI-generated embodied virtual teachers, this study intended to investigate the effects, application contexts, research methodologies, and detailed components of embodied virtual teachers generated by AI. We would systematically review publications to evaluate such technologies and provide a comprehensive view of current research. The following research questions were proposed based on the existing reviews, publications on this topic, and the review frameworks (e.g., Dai et al., 2022; Davis et al., 2021; 2023):

We searched four literature databases for the relevant publications on 6 June 2024. The keywords selected included (1) “artificial intelligence” OR AI and (2) virtual teacher* OR virtual instructor* OR teach* agent* OR teach* assistant* OR avatar teacher* OR embodied pedagogical agent*. Within each line, the alternative expressions were connected with a Boolean operator “OR”, and those two lines were connected with an “AND” to match literature records that contained AI technologies and reflected the application of virtual teachers. The proposed search strategies were applied to the Web of Science Core Collection, SCOPUS, Sage, and Springer Link. Asterisks were added when the database allowed the wildcards since they located more records that contained variations of similar roots. For example, “teach*” matched “teacher,” “teachers,” “teaching,” and “teachable.” In the Web of Science, we applied the strategies to the “Topic”, i.e., article titles, keywords, and abstracts. In SCOPUS, the strategies were applied to “Title-Abstract-Keyword”. The strategies were applied to “anywhere” for the other two databases because the restrictions on search fields would lead to few records. In Springer Link, we only included “conference paper,” “article,” and “chapter” since the database also collected many books and references that could not be included in the literature synthesis.

Two researchers independently filtered the search results based on four steps. First, we removed duplicate, incomplete, and retracted records with an automated literature management program, Zotero. Then, we evaluated the relevance of the search results. Pedagogical agent generation methods and tools were particularly examined to ensure that virtual teachers in the included studies involved AI generation. In the third step, we followed standards for reporting empirical research in social sciences (Duran et al., 2006) to assess the quality of empirical studies. We excluded empirical studies that (1) were not rigorously designed, (2) did not contain reliable findings for further synthesis, (3) were not empirical studies nor reviews, such as editorials and book reviews, or (4) did not have complete structures for academic publications. We only included studies that were rigorously designed, contained reliable results, and were structured in complete and academic logic. Although the standards primarily applied to empirical studies, the researchers proposed similar criteria for review articles. We only included review articles that (1) were rigorously conducted, (2) were comprehensive, (3) relied on large size of publications (e.g., included studies no less than ten), and (4) were reported in complete academic structures.

Disagreements between raters might occur about whether certain publications were eligible based on relevance and quality. As such, we invited a third rater to make ultimate decisions when the two raters did not agree on some filtration results. The three researchers included one distinguished professor specializing in technology-enhanced education research for more than five years and published hundreds of academic articles related to the research interest. The others were two postgraduates in similar research domains who had received systematic methodological training from the professor and had related publishing experiences in educational technologies. The inter-rater assessment was communicated online or in person in a university setting. To further ensure the reliability of the assessment, we calculated the inter-rater reliability to assess the rating and selection processes. The literature selection procedures ended with an acceptable inter-rater reliability (κ = 0.86). After relevance and quality assessments and the rating reliability evaluation, we sought accessibility of the publications, ensuring the included studies could be read for their full texts based on our resources. The subsequent coding and analysis relied on publications included throughout the selection process.

Two trained coders independently retrieved the following items from the included publications: (1) authors and publication years, (2) publication titles, (3) publication types, (4) learning content where AI-generated embodied virtual teachers were applied, (5) educational levels where the studies were conducted, (6) research participants (including the sample sizes and their identities), (7) the impacts of embodied virtual teachers on learning effectiveness, (8) research methods, and (9) research subthemes from the perspective of embodied virtual teachers’ components. Items (1) to (3) were exported from the literature databases. To further enhance the coding consistency, both coders followed Hsu et al.’s article (2012) to classify participants’ educational levels. The components of embodied virtual teachers could include physical characteristics like voice, gestures, facial expressions, and interactive gaze (Mayer and DaPra, 2012), following which both coders refined the subthemes of the included studies. The two coders compared and merged their coding results to form the final ones.

Most included studies yielded quantitative results from subjective questionnaire surveys and objective test scores. The System Usability Scale assessed learners’ perceived usefulness and ease of use when they participated in virtual learning with AI-empowered avatar teachers (Grivokostopoulou et al., 2020). The Agent Persona Instrument measured learner trust in virtual teachers and perceived embodiment (Pataranutaporn et al., 2022). Learning motivation, satisfaction, social presence, cognitive load, and attitudes were also measured with established scales (Deng et al., 2022; Li et al., 2016). Learners’ perceptions of goal achievement in conversations could also be rated according to their self-report (Nagendran et al., 2022). Learning gains in knowledge acquisition, transfer, and application were widely assessed by tests, such as multiple-choice and open-ended questions (Ba et al., 2021). Behavioral experiments like eye-tracking and word recognition also produced quantitative results (Pi et al., 2022; Saadatzi et al., 2018). Researchers explored learners’ attention and cognitive load with neurological approaches like electroencephalography (EEG), event-related brain potentials (ERPs), and functional near-infrared spectroscopy (fNIRS) (Peng et al., 2020).

Interviews revealed learners’ perceptions of AI-generated embodied virtual teachers. Vallis et al. (2024) and Seo et al. (2021) relied on interviews to elicit learner perspectives of interaction, presence, and embodiment in virtual learning with AI-generated teachers. Presented with virtual teachers, students felt less invasive to their privacy and more comfortable interacting with their teachers (Seo et al., 2021). However, learners felt uncomfortable with student-teacher interaction when AI enabled teachers to monitor their unconscious facial expressions, hence the risk of pretense and distrust (Seo et al., 2021). Close interactions between virtual teachers and students were not entirely negative. Students might be satisfied with their effective social interaction and critical roles in classrooms (Vallis et al., 2024). Regardless of the research methods, the included studies mostly investigated embodied virtual teachers generated by AI from the students’ perspectives (e.g., Schroeder et al., 2023). Among the included studies, teachers’ perceptions were found in Seo et al.’s semi-structured interviews involving higher-education instructors (2021).

Most studies explored the overall effectiveness of AI-generated embodied virtual teachers, while others concentrated on specific components, such as voices and appearances. Presenting a virtual human instructor in video-based learning, Craig and Schroeder (2017) found that computer-generated voices by modern software performed best in retaining students’ learning gains and inspiring knowledge transfer, which was confirmed by Pi et al. (2022). AI-generated voices also reduced the distraction of virtual instructor images and encouraged students to concentrate more on learning materials (Pi et al., 2022). Vocabulary learning effectiveness measured by memory test outcomes could be enhanced with iconic gestures of virtual instructors (Bergmann and Macedonia, 2013). As Li et al. (2016) suggested, appearance details indicated students’ liking, perceived social presence, and overall learning experiences with AI-generated embodied virtual teachers. Due to the significance of virtual instructors’ gestures and facial expressions, more recent designers adopted these human-like characteristics as indicators of their AI-generated embodied virtual teachers’ design and implementation (Chen et al., 2023; Li et al., 2023).

AI efficiently generates embodied virtual teachers based on pictures, texts, and big data of human faces, incorporates dynamic characters into videos, and improves the quality of virtual characters and generated videos, which aligns with the previous studies (Pellas, 2023). The efficiency of virtual image generation and the perceived realism distinguish AI-generated embodied virtual teachers from traditional methods, such as 2D, static, and animated agents. Some included studies consistently and successfully incorporate AI-generated embodied virtual teachers into videos for online courses and mobile applications (Shu and Gu, 2023). Such integration is greatly motivated by the combined advantages of real teacher presence and virtual characters in educational contexts. For the former, Yu (2022) established significant improvements in learning outcomes as a consequence of human teacher presence in videos; for the latter, virtual pedagogical agents improved learning achievements in mathematics and self-regulating skills (Mohammadhasani et al., 2018; Yılmaz et al., 2018).

Longitudinally, the included studies demonstrate a revolutionary technological advancement in virtual character generation. Previous AI technologies enabled researchers to create virtual teachers primarily through 3D modeling and rendering (e.g., Peng et al., 2020). Traditional characters relied on pre-designed components usually selected by the users (Hanna and Kim, 2021). In contrast, AI empowers prompt simulation and generation of virtual teacher presence with algorithms, with dominant applications in producing virtual talking-head videos (Zhen et al., 2023). Recent works have shifted to automated creations, and humans’ roles and workload in creating and implementing virtual teacher presence are further diminished (Pataranutaporn et al., 2021). Therefore, AI dynamically, synchronously, and flexibly generates embodied virtual teachers with prepared elements. Researchers probed into fine-grained components of embodied virtual teachers to achieve better presentations (e.g., Li et al., 2023), consistent with previous reviews on pedagogical agents (Dai et al., 2022). AI-generated teachers also allowed researchers to incorporate more recent technologies, such as natural language recognition and synchronized fine movements (Pi et al., 2022; Shu and Gu, 2023).

Virtual pedagogical agents play critical roles in virtual learning environments when AI empowers them with more than vividly generated images. Inconsistencies existed regarding the effectiveness of pedagogical agents (e.g., Davis et al., 2021), which is also supported by this review when both positive and negative effects are found: The included studies add up to mostly positive impacts of AI-generated embodied virtual teacher presence by measuring the experiential aspects and performances. The effectiveness measurements in the included studies echo some extended Bloom’s taxonomies of learning objectives, i.e., cognitive, psychomotor, and affective domains (Zhou and Brown, 2015). When the roles of generated embodied virtual teachers are examined, this review concludes with their dominant role in delivering learning content (e.g., Shu and Gu, 2023). We concentrate on a particular category when Dai et al. (2022) classified pedagogical agent characters into experts, motivators, mentors, and learners. However, students perceived their power in additional functionalities in virtual learning environments, including (1) balancing social connections and privacy, (2) facilitating educational needs analysis, and (3) analyzing fine cues in students’ facial expressions and body language captured by AI (Seo et al., 2021).

Refined investigations of virtual teacher presence in this review do not indicate more professionalized and specialized roles of avatar teachers in empirical studies. Nevertheless, AI-generated embodied virtual teachers encouraged researchers to incorporate different AI technologies and synthesize their advantages (e.g., Guo and Gao, 2022). This demonstrates the powerful compatibility of AI technologies in education that empower and strengthen relatively traditional ones. For another example, natural language recognition and machine learning techniques empowered virtual learning environments (Rivas et al., 2021). Also, different from the overall view of technological implementations of pedagogical agents (Dai et al., 2022), virtual reality and metaverse became even more important than the traditional avatar teacher presence (e.g., Grivokostopoulou et al., 2020). With the introduction of AI-generated embodied virtual teachers, students interact differently with their peers, instructors, and educational technologies.

Student-teacher relationships are reshaped, given that students may more closely interact with their instructors in the diminished power distance. The transformative presence of virtual teachers raises the same question, “where are the educators?”, proposed in Zawacki-Richter et al.’s systematic review (2019, p. 39) after they examined the roles of AI in education. Virtual teachers have taken the primary roles in delivering lectures in the included studies, regardless of the learning content, and students perceived the flexibility of the virtual teachers (e.g., Daniels and Lee, 2022). As traditional lecturers, motivators, and question breakers, human teachers are exposed to the risk of being replaced because many included studies in this review have reported enhanced learning outcomes in experiences, cognition, and behavior with AI-generated embodied virtual teachers. AI-generated embodied virtual teachers are breaking the hierarchical differences between teachers and students, attracting learners’ interest as a novel technology in educational contexts. With the compatibility with other AI technologies, they develop constantly growing functionalities, resulting in the threatening risk of replacing human teachers.

On the other hand, as teaching includes much more than delivering the learning content, the indispensable roles of human instructors are particularly noticeable. For one point, the enhancement of learning outcomes and experiences is unstable. Teachers and researchers need continued attention in identifying what educational technologies yield the best outcomes in diverse educational levels, learning content, and participants through scientific research and teaching practice. Revealing such intricate mechanisms depends on creative reasoning and human experience. The negative and insignificant impacts inspire human instructors to seek more suitable technologies in specific contexts. Researchers and instructors should also be reminded of the usability and implications of insignificant findings in educational domains, according to Edelsbrunner and Thurn’s (2024) review. In this sense, human instructors will be coordinators of multiple educational technologies in the future, in addition to giving lectures, designing technology-enriched curricula, and training students to use technologies. As task-technology fit theories advocate, the matching relationships between task and technology characteristics influence learner performance in technology-assisted learning (e.g., Al-Rahmi et al., 2023).

In another aspect, with Li et al.’s explorations into human-AI collaboration modes (2023), human efforts will still be valuable in exploring and adapting educational resources, especially under specific educational policies and socio-cultural contexts. Most AI-generated embodied virtual teachers only simulated human instructors or presented what human instructors programmed them to teach. Exploring educational resources and preparing students with the ability to harness AI technologies will rely mostly on human instructors. This is consistent with recent studies in AI-empowered learning technologies. Reflecting on personalized recommender systems, Brusilovsky established the significance of transparency and controllability for collaborative recommendation; meanwhile, researchers also found that young learners seemed to have insufficient knowledge of the learning domains to navigate the entire learning process (Brusilovsky, 2023). Including the ethical principles to deal with the technology, data, and AI-generated results, students’ AI literacy should primarily be shaped through human instructions.

Ethical concerns about AI-generated embodied virtual teachers encompass multiple areas. The included studies tested students’ perceived credibility and acceptance of the generated virtual teachers following the perceived embodiment and interpersonal attraction (e.g., Li et al., 2016). However, the degree of realism and embodiment also brought difficulties distinguishing generated avatars from real persons, causing the risk of manipulation and control (Pasquale and Selwyn, 2023; Vaccari and Chadwick, 2020). Students’ credibility in AI technologies will be possibly harmed since they aim to learn knowledge rather than gain pure entertainment. A familiarization process with virtual teachers soon eroded students’ curiosity and excitement (Vallis et al., 2024). This process wiped out critical motivating factors like social presence and interpersonal attraction for accepting and adopting technologies with virtual teachers (Huang et al., 2022). Another concern involves exploiting learner data collected by AI-empowered systems for educational purposes. Privacy and data security may explain students’ distrust and concerns about AI-generated embodied virtual teachers.

Coming along with flexible and distributed learning are concerns about the accessibility, sustainability, and educational equity of learning tools with AI-generated embodied virtual teachers. This is congruent with Li et al.’s argument about instructional videos (2016) that many learners have difficulty accessing digital learning resources, although most technologies aim to enhance educational equity and openness. Vallis et al. (2024) also suggested the potential of AI for providing open educational resources. However, rapid advancements in educational technologies may pose other challenges, including students’ and teachers’ digital literacy in the AI era, financial support for smart learning environments, and sustainable utilization of educational technologies. The concern is not unique to the included studies. Metaverse that allowed AI-generated virtual presence contributed to open educational resources (Zhai et al., 2023). However, if learners have unequal access to AI-empowered learning materials, unintended gaps will be opened with such technologies. Concerns about accessibility, sustainability, and educational equity with AI-generated embodied virtual teachers are only parts of the overview of AI in education research. Common challenges have been proposed and alerted by many other researchers (e.g., Schiff, 2021).

Promoting learning effectiveness with embodiment and virtual presence contains a multi-faceted mechanism, and the long-term effect of embodied virtual teachers is not entirely crystallized. As Dai et al. indicated (2022), many contextual factors influenced the effects of pedagogical agents. The included studies also presented insignificant effects of embodied virtual teachers compared with human instructors and static teaching agents (e.g., Bergmann and Macedonia, 2013). Learning content is noticeable when the included studies have extensively reported the applications in natural sciences and language learning. In contrast, applications in other subjects need extended efforts. Another critical aspect is learning media to present virtual teachers, forming diverse virtual learning contexts in the metaverse, virtual reality, and computer-mediated learning (Pataranutaporn et al., 2022; Shu and Gu, 2023). Like embodied virtual teachers, many educational technologies may facilitate learning by enhancing social presence, interaction, emotional support, and learning effectiveness (Parmaxi, 2023; Yu, 2022). Similar to what we discussed above in teachers’ reshaped roles, this concern should remind researchers and instructors of task-technology fit in examining technology acceptance and adoption (Goodhue and Thompson, 1995).

This systematic review concentrates on AI-generated embodied virtual teachers. Based on 31 included empirical studies, this review analyzes the effects, application contexts, research methods, and design features of AI-generated embodied virtual teachers. We find that such virtual teachers demonstrate mostly positive impacts on learning measured by experiential aspects and performances despite some insignificant or negative results. Additionally, they have been dominantly applied to higher education and teaching natural science subjects. Corresponding to the effectiveness measurements of two primary categories, researchers have adopted questionnaire surveys based on multiple scales for the related variables, interviews, and standardized tests for data collection. Fine components of the embodied virtual teachers have been investigated to examine the effects on learning outcomes, including facial expressions and voices as two dominant parts. Our findings contribute to a better understanding of new features of AI-generated embodied virtual teachers, the roles of human and virtual teachers, and concerns about their applications.

This study contains two limitations. First, unlike quantitative synthesis approaches like meta-analysis, the effects of AI-generated embodied virtual teachers were not quantified to gain statistically stable results. However, this study could still establish its contributions with emerging topics and findings from an innovative focus on AI empowerment. It can also extend the existing findings that did not particularly concentrate on AI. Second, this study did not interpret the findings from a technological perspective that is beyond our knowledge scope. Our findings can be enlightening and representative among educational studies, but future research is still encouraged to extend our findings to generative AI algorithms and tools based on expertise in related fields.

Our review has both theoretical and practical implications. Theoretically, AI-generated embodied virtual teachers stand at the intersection of multiple branches of educational research and theories, including social presence, human-computer interaction, and experiential learning theory. Our review provides an example of practicing our comprehensive and integrated review framework in coding the reviewed articles and discussing emerging topics. Students’ social presence has been closely associated with the perceived embodiment of teacher presence in virtual learning environments (Li et al., 2016; Suk and Laine, 2023), indicating a new way to enhance students’ social presence in technology-assisted learning environments. Empirical studies have established frameworks of embodiment and embodied cognition (Walkington et al., 2023; Wilson, 2002). The design practice in AI-generated virtual embodied teachers includes concrete examples of these frameworks. Under the framework of human-AI collaboration in education, AI-generated embodied virtual teachers are less investigated, while this technology can dominate the role of delivering learning content and engaging students in classes. The findings of this study shed light on human instructors’ roles that have been drastically restructured, encouraging teacher education to seek proactive exploration and transformation. Future designers can continue to enrich the related theories and explore fine components of embodied designs and highly effective designs of embodied virtual teacher presence.

Practically, this review encourages instructors and researchers to implement this technology further. AI-generated embodied virtual teachers may be improved to fit more diverse learning contexts. Alternatively, combining diverse educational technologies may solve the dilemma caused by the insignificant impacts of AI-generated embodied virtual teachers in specific contexts. Technology acceptance research is still needed for this technology to evaluate the fitness between tasks, technologies, and persons. For instance, Huang et al. (2022) have examined machine teachers as a composite concept and established the impacts of interpersonal attraction and social presence on students’ acceptance and adoption of machine teachers. The impacts of perceived embodiments of AI-generated embodied virtual teachers on technology acceptance indicators have been established, with perceived interactivity and learner impact as significant moderators (Lin and Yu, 2025). Wang et al. (2024b) established the critical role of AI literacy in determining learner acceptance and adoption of generative AI, highlighting its increasingly important role in quality and future-oriented education. Future research can extend such methods to embodied virtual teachers generated by AI, probing into more influencing factors.

Additionally, most studies relied on cross-sectional methods, which leaves the long-term effects unclear, especially with many unsolved ethical concerns. From the perspective of firms and technological designers, virtual character production should be explicitly marked as “AI-generated” to avoid misunderstanding and misinterpretation of videos, as Lin and Yu suggested (2025). Evidence and innovative practice from more diverse contexts may help address the contradictions. Meta-analytical evidence finds the overall effects of pedagogical agents, while more empirical studies can be conducted to explore the effects of such technologies to address some inconsistencies in the effectiveness. With the significantly positive evidence regarding the components of AI-generated embodied virtual teachers, firms may yield the influences of specific designs, such as appearances, voices, and gestures, to attract future learners and enhance their learning outcomes.

This review also indicates a dearth of implementation at pre-college levels. More empirical evidence should be encouraged at these levels, like in Deng et al. (2022). Similarly, the uneven popularity of different research methodologies and learning content can be identified. Teachers’ perceptions about this technology are to be clarified. Future research can contribute to less investigated application contexts and populations. This technology can be implemented in particular pedagogical designs, for instance, to facilitate personalized learning (Zhai et al., 2023). Designers and instructors are encouraged to develop content-specific virtual teachers to facilitate learning in various contexts. For example, language learning applications on mobiles have just started using virtual characters as learning partners to interest their users (Wan and Moorhouse, 2024). The subject-specific virtual characters and learning materials demonstrate efficient and enlightening practices of this technology. Future instructors should also actively and innovatively explore the transformative pedagogical patterns and seek collaboration with AI (Järvelä et al., 2023) when AI-generated embodied virtual teachers are to dominate the teaching roles. When teaching is partially replaced by virtual avatar teachers and student-teacher interaction is reshaped, the consequences on pedagogical approaches and the irreplaceable roles of human teachers should be open to innovative explorations.