AI in Education from CAI to AGI: A Narrative Review and Conceptual Framework

1. Introduction

With the rapid development of information technology, Artificial Intelligence (AI) technology is profoundly transforming the landscape of the education sector and driving revolutionary changes in educational and teaching methodologies. AI has not only elevated the standard of educational technology, but also reshaped the core of educational models, refined teaching methods and learning experiences, and provided robust technical support for the innovative development of education (Walter, 2024; Chen et al., 2020). Driven by AI technology, education is undergoing a transformative upgrade toward digitalization (Carmo et al., 2025), personalization (Strielkowski et al., 2025; Liu, 2025), and intellectualization (Xie and Jiao, 2025). From early computer-assisted instruction and online learning platforms to contemporary AI-supported personalized learning systems, technological advances have continuously reshaped the organization of teaching and learning. In particular, the recent development of generative AI has introduced new possibilities for interactive learning support, content generation, and human–AI collaboration, while also raising new pedagogical and ethical concerns (Chiu, 2024; Šedlbauer et al., 2024; Giannakos et al., 2025).

Recent advances in large language models (LLMs), multimodal foundation models, and other general-purpose AI technologies have renewed discussion about the longer-term possibility of Artificial General Intelligence (AGI) (Raman et al., 2025; Fei et al., 2022). AGI possesses cross-domain (Raman et al., 2025) and multimodal (Fei et al., 2022) cognitive capabilities. It can not only comprehend and learn knowledge, but also flexibly transfer and apply such knowledge in complex tasks, demonstrating an intelligence level close to that of humans. At present, it is generative AI and other advanced AI technologies that are being integrated into educational practice to support personalized learning, content generation, interactive assistance, and flexible instructional experiences (Gu et al., 2023). At the same time, their growing use has raised a range of ethical and educational concerns, including the digital divide, privacy protection, algorithmic fairness, and the changing relationship between teachers and intelligent systems (Wang et al., 2024; Al-Zahrani and Alasmari, 2024; Simon, 2019; Chan and Tsi, 2024). Against this background, how to scientifically and rationally integrate AI into teaching practice, balance the role relationship between human educators and intelligent systems, and avoid the blind application or even abuse of technology has become a critical issue that needs to be addressed urgently (Chan and Tsi, 2024). Therefore, it is necessary to systematically sort out the inherent logic of technology-driven evolution of teaching models, clarify its advantages and limitations, and provide theoretical support for the innovation of educational technology and the improvement of relevant systems.

To address the aforementioned issues, this study traces the evolution of AI-enabled education and summarizes the corresponding transformation of educational and teaching models. Focusing on key technologies such as intelligent teaching assistants, extended reality (XR) technologies, and large language models (LLMs), this paper examines their practical applications and analyzes how different forms of AI support innovation in teaching and learning. Building on these findings, the review proposes an evidence-informed Technology–Mechanism–Outcome–Challenge framework that positions learning mechanisms as an intermediate explanatory layer between technological affordances and educational outcomes. It further identifies teacher-role transformation, learner overreliance, digital inequality, privacy, transparency, bias, and human oversight as cross-cutting challenges. Looking toward the prospective development of AGI, the review argues that broader and more autonomous AI capabilities may amplify these challenges, while learner agency, human oversight, equity, transparency, and responsible governance remain essential across technological stages.

The remainder of this paper is structured as follows: Section 2 describes the narrative review approach and literature selection strategy. Section 3 traces the evolution of AI-enabled educational and pedagogical models. Section 4 examines representative applications and evidence concerning intelligent tutoring systems, XR-enabled learning, and generative AI. Section 5 synthesizes these findings through a Technology–Mechanism–Outcome–Challenge framework and explains the mediating role of learning mechanisms in translating technological affordances into educational outcomes. Section 6 discusses pedagogical, equity, ethical, and governance challenges in current AI-enabled education and considers their implications for the prospective AGI horizon. Section 7 concludes the review and outlines future directions.

2. Review Approach and Literature Selection

2.1. Review Approach

This study adopts a narrative review and conceptual synthesis approach to examine the evolution of artificial intelligence in education from computer-assisted instruction (CAI) to the emerging prospect of AGI. Rather than aiming for exhaustive coverage of all publications, the literature corpus was selected purposively and illustratively to represent major technological paradigms, educational mechanisms, applications, and challenges across different stages of AI-enabled education.

The primary literature corpus focuses on studies published between 2010 and 2026. Earlier seminal works were additionally retained when necessary to establish the historical and conceptual foundations of CAI, intelligent tutoring systems (ITS), cognitive tutoring, and related educational paradigms. These earlier sources were used primarily for historical contextualization rather than as part of the contemporary evidence base.

2.2. Search Strategy

Relevant literature was identified primarily through Web of Science, Scopus, ERIC, IEEE Xplore, and Google Scholar. Searches combined terms related to artificial intelligence and education, including “artificial intelligence”, “computer-assisted instruction”, “intelligent tutoring system”, “adaptive learning”, “learning analytics”, “extended reality”, “virtual reality”, “augmented reality”, “generative AI”, “large language model”, “ChatGPT”, and “artificial general intelligence”, together with educational terms such as “education”, “teaching”, “learning”, “student”, “teacher”, and “higher education”.

Additional searches were conducted for themes related to educational impact and governance, including “academic integrity”, “critical thinking”, “privacy”, “fairness”, “algorithmic bias”, “digital divide”, and “teacher–AI collaboration”. Reference lists of relevant reviews and representative studies were also examined to identify additional foundational or influential publications.

2.3. Purposive Sampling and Selection Criteria

Literature selection followed a purposive and illustrative sampling strategy. Priority was given to studies that met the following criteria: (1) publication between 2010 and 2026, except for seminal historical works; (2) publication through established or indexed academic venues; (3) direct relevance to AI-enabled teaching, learning, educational assessment, teacher support, or educational governance; and (4) contribution to at least one major mechanism examined in this review, including intelligent tutoring and adaptive learning, data-driven learning analytics, XR-supported learning, generative AI and LLM-supported education, or human–AI collaboration.

When discussing educational effectiveness, greater emphasis was placed on systematic reviews, meta-analyses, controlled or longitudinal studies, and substantive empirical evaluations. Studies concerned only with general AI model performance, without a meaningful educational process, learner or teacher outcome, or pedagogical implication, were not treated as primary evidence of educational effectiveness. Conceptual papers, books, and policy reports were selectively retained when they contributed to historical interpretation, theoretical development, or ethical and governance discussions.

The final literature corpus comprised 104 references, including 40 primary empirical or system-evaluation studies and 64 contextual sources, such as reviews, meta-analyses, conceptual and theoretical works, books, and policy-oriented publications. Literature selection was conducted purposively by the author team to provide representative coverage of the technological paradigms, educational mechanisms, applications, and challenges examined in this review. A formal dual-reviewer independent screening procedure was not implemented, and inter-rater agreement was therefore not calculated. Consistent with the narrative and conceptual-synthesis design of the study, a PRISMA flow diagram is not reported. The review does not claim exhaustive identification or screening of all eligible publications, but instead aims to provide a transparent and evidence-informed synthesis of representative literature relevant to the scope of the study.

3. The Development Course of AI in Education and Teaching

The explosive development of AI is driving education and teaching into a critical transition period of paradigm reconstruction. To profoundly grasp the trajectory of this educational transformation, it is imperative to first clarify its core driving mechanism, systematically sort out the development context of the integration of AI with education and teaching, and present the underlying logic and full development landscape of this educational transformation.

The development of AI in education and teaching is a progressive evolutionary process from an “auxiliary tool” to a “cognitive partner”, and further to a “reshaper of the educational ecosystem”. Its development trajectory is not a linear leap; instead, it has formed stages with distinct characteristics under the intertwined driving forces of technological breakthroughs, the renewal of educational philosophies, and social demands. Taking the core technological paradigm and the depth of its integration with education as the primary basis, this development process can be divided into the following four critical stages, as shown in Figure 1.

3.1. The Period of AI Emergence

The 1950s to 1980s marked the embryonic stage of the educational application of AI. The essence of this stage lies in the initial integration of the concept of “programmed instruction” guided by behaviorist learning theory with early computer technology. At the technical level, this stage relied on mainframe computers and early personal computers to deliver teaching through pre-set programs and fixed algorithms. Typical forms included the early prototypes of CAI (Atkinson and Wilson, 1968) and Intelligent Tutoring Systems (ITS) (Nwana, 1990). Such systems broke down knowledge into linear or simply branched teaching sequences, which were essentially an extension of “electronic programmed textbooks”.

During the 1980s, educational applications of AI and computing began to diversify beyond conventional CAI. Educational expert systems increasingly applied knowledge-based reasoning to diagnosis, guidance, and tutoring, while the LOGO programming environment encouraged learners to explore, program, and construct knowledge through active interaction with computers. Papert’s work with LOGO laid important foundations for the constructionist tradition, which emphasized learning through the creation of personally meaningful artifacts rather than passive knowledge reception (Papert and Harel, 1991). These developments broadened the pedagogical role of computers and provided an important transition toward the more adaptive and learner-centered systems of the 1990s (Papert, 2020; Nwana, 1990; Wenger, 2014).

At the educational application level, the core objective of this stage was to realize preliminary individualized teaching and support learners to complete learning tasks at their own pace. However, its so-called “intelligence” was only reflected in rule-based path branching (e.g., switching learning units according to answering results), which failed to truly identify learners’ cognitive states and lacked dynamic modeling and adaptive adjustment of the learning process. During this period, AI could only play the role of a “tutoring machine” with limited functions. Although it could improve the efficiency of repetitive training, it had a single interaction mode and a rigid feedback mechanism. Furthermore, due to the high cost of hardware and systems, the widespread popularization of AI was ultimately not achieved.

3.2. The Period of Model-Driven AI Development

From the 1990s to the early 21st century, with the improvement of computer performance and the development of related disciplines such as cognitive science and educational psychology, the application of AI in education gradually entered a model-driven stage, the core development of which lay in learner modeling (Du Boulay et al., 2023). During this period, AI systems began to attempt to infer learners’ knowledge mastery levels, cognitive characteristics, and potential misconceptions by constructing learner models, thereby improving the pertinence and adaptability of teaching support. The key underpinnings of this stage mainly included cognitive modeling, Bayesian Knowledge Tracing, and preliminary machine learning methods. Based on these technologies, adaptive learning systems and ITS were further developed, with a typical representative being the early Cognitive Tutor (Anderson et al., 1995).

In the field of education, influenced by cognitivist learning theory during this period, greater attention was paid to the analysis and support of learners’ internal cognitive processes. Teaching activities no longer merely stayed at content presentation and behavioral feedback, but gradually shifted to a model-based “diagnosis-intervention” model. However, the model construction had poor scalability and limited support for open knowledge, and truly individualized teaching could not yet be realized.

3.3. The Period of Data-Driven AI Education

From 2010 to 2022, driven by the rapid development of technologies such as big data, cloud computing, and deep learning, the application of AI in education gradually entered the data-driven stage. The most prominent feature of this stage is that AI identifies learning patterns, predicts learning risks, and supports educational decision-making through the collection, processing, and analysis of large-scale learning behavior data (Ahmad et al., 2023; Salas-Martínez and Ramírez-Martinell, 2024). Relying on online learning platforms, systems can continuously collect a large volume of behavioral data generated by learners during digital learning processes, and implement applications including learning risk identification, dropout early warning, and resource recommendation on this basis. AI thus began to provide data-driven, precise services for large-scale learning scenarios in the field of education.

In the field of education, this stage was influenced by connectivism and learning science, with an emphasis on understanding learning patterns through data in authentic, large-scale learning environments. However, while the data-driven paradigm has brought about efficiency improvement and service optimization, it has also exposed a series of limitations, such as data privacy risks and algorithmic bias.

3.4. The Period of Generative and AI-Enabled Education

The rapid development of generative AI represented by LLMs marks that education and teaching have entered a new stage with generative interaction and cognitive enhancement as its core characteristics. Since 2022, applications based on LLMs such as ChatGPT have emerged one after another, marking that AI technology has made critical progress in advancing toward AGI (Gu et al., 2023). Generative AI in this stage can not only understand the questions raised by learners, but also generate explanations, examples, feedback, materials and tasks around specific scenarios, embodying stronger interactivity, generativity and collaboration.

Looking beyond current generative AI, the prospective emergence of AGI could further reshape the roles and relationships of stakeholders in the education system. For example, Xu et al. proposed a conceptual framework and practical scheme for the standardization system of AI models in the field of education, which clarifies the basic principles and application strategies for teachers and students to use such AI tools. They emphasize that teachers and students should correctly understand the value and limitations of AI, maintain a prudent attitude toward AI-generated content, and improve their literacy and competencies in collaborative work with AI (Xu et al., 2025). In this transformation process, the role of teachers will shift from knowledge imparters to learning guides, innovation inspirers, and personality shapers; the role of students will transform from knowledge consumers to generative co-creators; and AI will evolve from a tool to a human-like agent, jointly assuming educational responsibilities with teachers and students to provide learners with a more personalized, flexible, and inclusive educational experience.

In the field of education, the rapid development of generative AI and related AI technologies has contributed to important changes in learning-support paradigms. For example, the development of generative AI has added a new dimension to ITS, enabling them to accurately model students’ cognitive states, learning styles, and knowledge mastery levels, thereby providing personalized and adaptive learning and teaching pathways (Liu et al., 2024). Meanwhile, AI in the field of education is gradually realizing multimodal fusion, encompassing multiple data forms such as text, images, videos, and audios, to construct a more comprehensive learning analytics framework and provide more accurate and insightful support for educational decision-making (Lee et al., 2025). AI-based adaptive learning systems can continuously track students’ progress, analyze their strengths and weaknesses, monitor changes in their learning states, promptly identify learning difficulties, and adjust teaching content and progress in real time (Yenduri et al., 2025; Yolcu, 2023). Furthermore, Virtual Reality, Augmented Reality, and Mixed Reality technologies, empowered by generative AI, also deliver more immersive and interactive learning experiences for learners (Tracy and Spantidi, 2025; Chamola et al., 2024).

However, while the growing adoption of generative AI has created new opportunities for education, the education system is also confronted with numerous challenges, including the potential weakening of learners’ agency and cognitive development, the further widening of the digital divide, the increasingly prominent issues of educational ethics, and the risk of the weakening of humanistic values (Muthukrishna et al., 2025).

4. Practical Applications of AI-Enabled Education

With the rapid development and growing adoption of AI, particularly generative AI, educators and educational institutions have increasingly explored how these technologies can support teaching, learning, assessment, and instructional decision-making. This Section conducts case analysis and summary from the following three aspects: intelligent teaching assistants, XR empowerment, and LLMs.

4.1. Intelligent Teaching Assistants

ITS has become a critical tool in educational technology. Such systems model students’ cognitive states, learning styles, and knowledge mastery levels based on AI, so as to provide personalized learning pathways and real-time feedback (Bloom, 1984; VanLehn, 2011). By systematically reviewing the relevant literature on ITS, Suwita et al. concluded that three core components should be focused on in the design of ITS: the student model, the knowledge model, and the interaction model (Suwita et al., 2024). With the rapid development of generative AI in the field of education, ITS have also ushered in new technological transformations. For example, the study by Liu et al. integrated GPT-4 into ITSs, which significantly enhanced the effectiveness of the systems in personalized teaching and demonstrated the broad prospects of generative AI-enabled ITS (Liu et al., 2024).

Existing studies indicate that intelligent teaching assistants exert their functions primarily through mechanisms such as learner modeling, real-time feedback, adaptive learning pathway adjustment, and continuous learning companionship. These distinct mechanisms collectively form the fundamental framework for their learning support functions. Their main manifestations are presented in Table 1.

Whether through the dynamic identification of learners’ status, or via real-time prompts, pathway adjustment, and learning companionship, intelligent teaching assistants are driving the shift of learning support from static resource provision to dynamic process intervention. At the practical application level, existing studies suggest that intelligent tutoring and adaptive learning systems can support learning across different educational contexts, but the underlying evidence base is heterogeneous in terms of study design, sample size, and methodological rigor (Kulik and Fletcher, 2016; Ma et al., 2014). To provide a more transparent comparison of this evidence, Table 2 summarizes representative studies according to educational context, study design, sample or evidence base, learning mechanism and reported findings or outcomes.

4.2. Immersive Virtual Learning Environment

Unlike intelligent teaching assistants, which primarily function in cognitive diagnosis and feedback support, the advantages of XR technologies (including VR, AR and MR) in education are more predominantly reflected in situational construction, spatial representation, and practical training. By constructing highly realistic virtual experimental environments, reproducing historical scenarios, and presenting three-dimensional visual models of complex concepts, XR technologies provide students with a safe and controllable practical space. XR technologies also enhance the immersion, interactivity, and explorability of the learning process, and effectively facilitate the development of students’ deep learning and higher-order thinking skills. From the perspective of learning mechanisms, XR supports learning primarily through five approaches, as shown in Table 3.

In specific applied research, empirical evidence for XR-enabled education and teaching has gradually evolved from scattered case studies to systematic reviews, meta-analyses, and interdisciplinary comparative studies. Existing literature indicates that XR has not only been applied in fields emphasizing hands-on practice and spatial comprehension, including medicine, nursing, engineering, and STEM education, but has also been gradually extended to a broader range of conceptual learning, situational experience, and participatory activities. To more comprehensively present the supported tasks, learning mechanisms, outcome indicators, and evidence types of XR across diverse learning scenarios, this paper selects representative studies for systematic summarization, as shown in Table 4.

4.3. Generative AI–Enabled Learning Support

Generative artificial intelligence, particularly systems based on large language models (LLMs), has broadened the scope of AI-supported teaching and learning. Unlike earlier rule-based or predominantly predictive systems, generative AI can engage learners in open-ended dialogue, generate explanatory materials and examples, provide formative feedback, support problem solving and content creation, and assist teachers in instructional design (Salvador et al., 2025). These capabilities enable generative AI to function not merely as an information-delivery tool, but as an interactive component embedded in teaching and learning processes. It has already been applied across a range of disciplinary contexts, including programming education, language learning, and medical education (Chen et al., 2026; Wang, 2025; Moskovich and Rozani, 2025; Hui et al., 2025; Bhuyan et al., 2025).

In authentic educational settings, generative AI has been integrated into various learning activities. In programming education, it can support code generation, debugging, and feedback provision; in language learning, it can facilitate personalized practice and communicative training; and in medical education, it has been incorporated into problem-based learning and clinical training activities (Wang, 2025; Moskovich and Rozani, 2025; Garcia, 2025). These applications indicate that generative AI can support multiple stages of the learning process, including explanation, practice, feedback, and reflection. However, the educational value of such systems depends not only on their functional capabilities, but also on how they are integrated into instructional activities.

Alongside these application-oriented studies, another line of research has examined the task-level capabilities of LLMs in education-related domains. Comparative studies have evaluated model performance in anatomy question answering, medical licensing examinations, Python code generation, and language assessment (Singal and Goyal, 2025; Pattanshetti et al., 2025; Shahzad and Iqbal, 2025; Rasheed et al., 2024). These studies provide useful evidence for understanding the functional strengths and limitations of current models across different tasks. Nevertheless, specific model rankings should be interpreted with caution, as performance may be influenced by task design, evaluation criteria, model configuration, and the rapid iteration of model versions. More importantly, task-level performance does not directly demonstrate that the use of such systems leads to improved student learning outcomes.

Accordingly, recent research has begun to move beyond model capability benchmarking toward the evaluation of learning outcomes in authentic instructional settings. For example, Kestin et al. evaluated AI tutoring using a randomized controlled trial design, while Chen et al. investigated a teachable ChatGPT agent in programming education (Kestin et al., 2025; Chen et al., 2026). Such studies shift the research focus from whether an AI system can successfully perform an educational task to whether its integration into teaching and learning can lead to measurable learner development. Clarifying this distinction is important for assessing the practical educational value of generative AI and for guiding its effective integration into future instructional practice.

5. Technology–Mechanism–Outcome–Challenge Framework

5.1. Framework Components and Analytical Logic

Building on the historical analysis in Section 3 and the application evidence synthesized in Section 4, this review proposes the Technology–Mechanism–Outcome–Challenge framework shown in Figure 2. The framework consists of four analytically connected layers: Technology, Learning Mechanisms, Educational Outcomes, and Challenges.

A central proposition of the framework is that technological affordances do not directly determine educational outcomes. Instead, their educational value depends on the learning mechanisms activated through instructional design, learner interaction, and implementation context. Technology provides possibilities for learning support, whereas mechanisms explain how these possibilities are translated into changes in learning processes and outcomes.

The same technological affordance may therefore lead to different outcomes depending on how it is used. For example, a generative AI system may provide direct answers, scaffolded hints, explanatory dialogue, or reflective prompts. Although these functions rely on similar underlying generative capabilities, they may activate different levels of learner engagement and cognitive processing. Accordingly, the framework treats the technology–mechanism relationship, rather than technology alone, as the key analytical link between technical capability and educational value.

5.2. Interpreting Educational Evidence Through the Framework

Applying the framework to the evidence reviewed in Section 4 reveals different levels of maturity across the three technology–mechanism–outcome pathways. Intelligent tutoring and adaptive learning systems mainly operate through learner diagnosis, adaptive feedback, and pathway adjustment to support personalization and learning performance, although the evidence remains heterogeneous. XR-enabled learning shows a comparatively more consolidated pathway from immersive and situated mechanisms to engagement and learning outcomes, supported by a relatively stronger body of systematic reviews and meta-analyses. By contrast, generative AI provides increasingly visible forms of cognitive support and scaffolding, but its effects on measurable educational outcomes remain less firmly established.

The framework also explains why technical capability benchmarks cannot be treated as direct evidence of educational effectiveness. Demonstrating that an AI system can answer questions, generate code, or evaluate text establishes technological capability, but does not show whether that capability activates a productive learning mechanism or improves learner outcomes. Educational effectiveness therefore requires evidence connecting technological affordances, learning mechanisms, and measured outcomes.

6. Challenges in AI-Enabled Education

The effective application of educational technology is inseparable from the in-depth integration of educational concepts and practice, and requires comprehensive consideration of multiple factors including teacher transformation, student competency development, and ethics (Nguyen et al., 2023; Baker and Hawn, 2022; Shahzad et al., 2025; Francis et al., 2025; Holmes and Tuomi, 2022). If future AI systems acquire broader and more general capabilities, these studies still provide an important basis for anticipating how educational risks and governance requirements may evolve.

6.1. Teacher Role Transformation and the Demand for Integrated Platform Construction

As AI has gradually penetrated into core teaching links including knowledge transmission, skill demonstration, and learning diagnosis, teachers are facing unprecedented pressure of role transformation and challenges of competency reconstruction. While AI can effectively assist teaching in aspects such as curriculum planning, writing guidance, and outcome feedback, its widespread application may also trigger problems including academic misconduct and impeded development of students’ thinking and comprehensive competencies. Therefore, teachers not only need to proficiently master and appropriately apply AI tools, but also should adhere to the essential goals of education, and cultivate students’ correct values, critical thinking, and autonomous decision-making abilities (Samala et al., 2025; Qadir, 2023; Ahmed et al., 2024). Within increasingly automated educational environments, teachers’ pedagogical judgment, interpersonal interactions, contextual understanding, and ethical responsibilities retain irreplaceable importance (Chan and Tsi, 2024; Selwyn, 2019).

Jinhee Kim conducted a study focusing on Teacher-AI Collaboration (TAC) through in-depth interviews with 20 leading Chinese teachers in AI-enabled education. For the first time, the study systematically defined the core dimensions and correlational logic of the effective implementation of TAC, which provides a critical reference for instructional design and teacher professional development (Kim, 2024). Wang et al. proposed Tutor CoPilot, a human–AI approach designed to provide tutors with real-time instructional support and to scale access to expert tutoring practices (Wang et al., 2024). In response to the demand for the transformation of teachers’ competency structure brought by the rapid evolution of technology, relevant scholars have successively constructed a variety of teacher profiling frameworks. For example, Hu et al. proposed a teacher profiling framework consisting of three dimensions and eight sub-dimensions, namely teacher cognition and emotion, teacher knowledge and skills, and the interaction between teacher cognition and competency. They also identified the potential challenges of the framework, which provides an important reference for the development and transformation of teachers’ professional competencies (Hu et al., 2024).

In addition to competency reconstruction at the individual teacher level, another key bottleneck currently restricting the deep integration of AI into teaching practice lies in the lack of a user-friendly, fully functional integrated platform. Functions such as curriculum design, resource generation, assignment grading, learning diagnosis and feedback are often scattered across different tools. Teachers are required to frequently switch between multiple tools, import and export data, repeatedly set prompts, and continuously conduct result verification and error correction. This situation confines the application of AI to fragmented use, making it difficult to form systematic teaching practices. Therefore, there is an urgent need to develop an implementable integrated AI teaching platform for teachers. Meanwhile, the platform should be embedded with compliance mechanisms to alleviate the burden on teachers in terms of quality control and risk avoidance.

6.2. Risks of Alienation of Students’ Learning Modes and the Demand for Academic Integrity Governance

Against the backdrop of the rapid popularization of intelligent tools, as users of AI tools, students are undergoing structural changes in their learning modes. Generative AI can provide nearly “instant and complete” solutions for tasks including problem solving, writing, programming, and literature integration. While this significantly improves learning efficiency, it also induces the substitution of students’ learning behaviors. Students’ attention has shifted from concept construction and problem representation to result acquisition and expression packaging, and the time for independent thinking has been significantly compressed. Over time, students’ metacognitive ability and reflective awareness will decline, and the learning process will be locked in the information comfort zone constructed by AI, leading to a trend of degradation in critical thinking and autonomous decision-making abilities (Singh et al., 2025; Bastani et al., 2025).

In addition, students may avoid learning input through methods such as AI ghostwriting, AI-completed assignments, and AI surrogate learning, disguising process-based learning as outcome-based output, thus forming a covert cheating chain (Cotton et al., 2024; Eke, 2023). Existing studies have pointed out that while generative AI brings opportunities for learning and scientific research, it is also accompanied by threats to the learning process, academic integrity, and competency development, which requires the education system to respond from the strategic and governance levels (Francis et al., 2025; Rahman and Watanobe, 2023).

The core of education lies not only in knowledge acquisition, but also in the shaping of a sound personality, the formation of value judgment ability, and the cultivation of social responsibility awareness. The education system needs to make simultaneous efforts at the levels of institutional norms, curriculum design, and learning support, to guide students to position AI as a cognitive empowerment tool rather than a cognitive substitution tool, and avoid the long-term impairment of key competencies such as critical thinking and innovation ability caused by the alienation of learning modes (Nie et al., 2025).

6.3. Ethical Challenges in the Use of AI

6.3.1. Digital Divide

The digital divide has long been a key bottleneck for countries worldwide in advancing the digital and intelligent transformation of education. With the rapid development of AI, the digital divide in the traditional sense is gradually evolving into a more complex multi-dimensional divide (Antoninis et al., 2023). The National Educational Technology Plan released by the United States in 2024 puts forward initiatives from three dimensions: the digital access divide, the digital design divide, and the digital use divide. In the context of higher education, Wang et al. took four higher education institutions in China that apply AI technology as research objects, and verified through modeling analysis that the AI capability of higher education institutions consists of three dimensions: resources, skills, and awareness (Wang et al., 2023).

Table 5 not only clearly presents the multi-dimensional structure of the digital divide, but also reveals the formation mechanism and governance priorities of each dimension, which provides a decision-making reference for education administrators, schools, and policymakers.

6.3.2. Data Privacy and Security

In the context of educational applications of AI, data privacy and security serve as the prerequisite and fundamental guarantee for the implementation of the technology, the core of which lies in the protection of sensitive data throughout the entire education process. Applications including intelligent assisted teaching, learning situation analysis, and personalized learning all require in-depth processing and analysis of massive, multi-dimensional educational data. Therefore, during the processes of data collection, transmission and use, it is imperative to prevent data from being illegally acquired, misused, leaked, or tampered with, so as to effectively protect students’ privacy and data security (Rawas, 2024; Yang et al., 2025; Al-Zahrani, 2024). Within the European context, the GDPR (The European Parliament and the Council of the European Union, 2016) is the main legal framework for regulating personal data processing. Its principles are of direct reference value for educational AI systems handling student data. Currently, existing studies have proposed combining federated learning and differential privacy technologies to build privacy-preserving AI models, enabling institutions to conduct intelligent and scalable learning analytics without crossing legal boundaries or relinquishing control over their own data governance (Imohimi, 2025; Desai et al., 2026). However, technical safeguards alone cannot substitute for legal and institutional accountability. Privacy-preserving educational AI therefore requires the integration of technical protection, data-governance procedures, and compliance with applicable data-protection requirements.

6.3.3. Algorithmic Transparency

Algorithmic transparency is a key prerequisite for avoiding implicit bias and enhancing the trust of the academic community and educational practitioners in AI (Marín et al., 2025). In the application scenarios of AI in education, algorithmic transparency enables teachers to more clearly understand the decision-making process of intelligent systems, thereby improving the perceptibility and controllability of the teaching process. However, the practice of transparency faces dual tensions: on the one hand, excessive emphasis on transparency may lead to the leakage of enterprises’ core technologies; on the other hand, overprotection of technical confidentiality may trap algorithms in a “black box” state, undermining the right to know and supervisory capacity of educational stakeholders. Therefore, it is necessary to explore a reasonable implementation path for algorithmic transparency that balances the protection of trade secrets and educational responsibilities. The EU Artificial Intelligence Act (The European Parliament and the Council of the European Union, 2024) provides a more concrete regulatory context for these concerns.

6.3.4. Algorithmic Bias

Algorithmic bias continuously undermines educational equity and has become a key bottleneck for the high-quality development of AI-enabled education (Yin et al., 2024; Kurumayya, 2025). Yin et al. pointed out in FairAIED that algorithmic bias in educational scenarios is not caused by a single link, but a systemic risk that runs through the entire chain of “data-algorithm-user interaction-institutional governance” (Yin et al., 2024). If the training data is biased in terms of population, culture, or other dimensions, the output of the model may present unfair or even discriminatory results. This will not only directly affect students’ learning experience and academic performance, but may further solidify educational disparities and violate the fundamental principle of educational equity (Samala et al., 2025). UNESCO’s (Miao and Holmes, 2023) human-centered guidance further reinforces the need to prevent AI adoption from reproducing or amplifying existing educational inequalities. Existing studies have conducted a comparative evaluation of three mainstream LLMs, namely ChatGPT, Gemini, and Claude, focusing on social bias and gender bias. The results show that ChatGPT has persistent bias, Gemini has occasional bias, while Claude has a relatively low level of bias (Shetty et al., 2024). This study reveals the significant differences in the degree of bias across LLMs, and further calls for the strengthening of governance strategies for AI bias to ensure the equity of AI in educational applications. In addition, the governance of algorithmic bias should extend the benchmarks to more challenging reasoning tasks and multimodal tasks, and establish a continuous measurement and fine-grained monitoring system that evolves over time, so as to construct verifiable and more stable fair AI systems in real educational and public service scenarios (Simpson et al., 2025).

7. Conclusions

This study reviews the evolutionary trajectory of AI in education and examines how technological advances have reshaped educational and pedagogical practices. The findings indicate that intelligent teaching assistants, XR technologies, and large language models are increasingly being integrated into key educational activities, including knowledge presentation, learning support, learning diagnosis, feedback, and instructional management. These developments reflect a broader transition from technology-assisted instruction toward more interactive, adaptive, and AI-supported forms of teaching and learning. The evidence-informed conceptual framework developed in this review integrates these findings by linking technological affordances with learning mechanisms and educational outcomes, while treating pedagogical, equity, ethical, and governance challenges as cross-cutting constraints rather than downstream consequences.

At the same time, the growing adoption of AI in education is accompanied by challenges related to educational equity, teaching quality, learner agency, data privacy, algorithmic transparency, and fairness. Addressing these issues requires not only continued technological development, but also appropriate pedagogical integration, institutional governance, and human oversight. Looking ahead, advances toward more general forms of artificial intelligence may further reshape educational models and stakeholder relationships. However, their educational value will depend on whether such systems can be integrated in ways that preserve human agency, promote equitable access, and support meaningful learning.