Supporting Self-Regulated Learning in the Age of Artificial Intelligence: A Comparative Systematic Review of Earlier AI and Generative AI

Abstract

Generative AI (GenAI) has rapidly emerged as a promising tool for supporting self-regulated learning (SRL). However, little is known about how its mechanisms compare with, or extend beyond, earlier artificial intelligence (AI) approaches. Using Winne and Hadwin’s COPES (Conditions, Operations, Products, Evaluations and Standards) architecture as an organizing framework, this comparative systematic review synthesized 70 articles from 2015 to 2025 to examine how AI and GenAI were technologically and pedagogically implemented to scaffold SRL. Technically, SRL systems predominantly employed knowledge-based systems, machine learning, natural language processing, and, more recently, customized large language models, with limited integration of external knowledge bases. Pedagogically, interventions in the reviewed studies concentrated on Operations and Evaluations during task enactment, with less attention to task definition, goal setting, and adaptation. In terms of operational distinctions between AI and GenAI, AI tended to instantiate analytics-driven scaffolding (data-based adaptivity, progress monitoring, and standards setting). In contrast, GenAI more often enabled dialogic, context-adaptive cognitive scaffolding and content generation. Building on these findings, this review proposes a framework for comprehensive SRL scaffolding.

Keywords

self-regulated learning artificial intelligence generative AI scaffolding metacognition the COPES model comparative systematic review

Introduction

The ultimate aim of education is to prepare individuals for lifelong learning; therefore, self-regulation is essential (Zimmerman, 2002). In this sense, self-regulated learning (SRL) contributes to the achievement of Sustainable Development Goal 4 (SDG 4)—Quality Education, which guarantees equity, inclusion, and lifelong learning opportunities for all people (United Nations General Assembly, 2015). SRL skills can be nurtured with explicit guidance and practice (Pintrich, 1995; Zimmerman, 2000), highlighting the potential of scaffolding to foster SRL development. Winne (2017) emphasized that technology, especially learning analytics and educational data, can promote effective SRL.

Artificial intelligence (AI) has been widely used throughout its development to support SRL through intelligent tutoring systems (ITS), chatbots, machine translation, and automated feedback tools (Chang & Sun, 2024). These technologies provide scaffolding for goals and plans, self-monitoring, and cognitive, metacognitive, and affective feedback (e.g., Han et al., 2026; Mejeh & Rehm, 2024). To date, AI technologies have remained important in SRL research, despite the recent surge in generative AI (GenAI).

Since ChatGPT’s release, GenAI has been widely employed to support SRL across disciplines (e.g., M. Liu & Reinders, 2025; Neumann et al., 2025). These interventions typically respond to learners’ requests for feedback and knowledge, and generate practice exercises (e.g., Huang & Lin, 2024; Neumann et al., 2025). In this context, an important question arises concerning how GenAI has been leveraged to support SRL, particularly from both technological and pedagogical perspectives. This question remains underexamined in previous reviews (e.g., Sardi et al., 2025).

Furthermore, although GenAI technically represents the next stage in the broader AI stream, existing systematic reviews have examined GenAI or AI mainly in isolation (e.g., Chang & Sun, 2024; Sardi et al., 2025). A recent mapping review by Banihashem et al. (2025) integrated GenAI into the broader AI stream to explore the AI–SRL relationship; however, GenAI-based interventions were not highlighted. As GenAI has rapidly emerged as a valuable SRL aid, we contend that a comparative analysis can clarify how GenAI and its mechanisms relate to, or extend beyond, those of earlier AI approaches. Additionally, although Banihashem et al. (2025) reported the most frequently used AI interventions, for instance, adaptive systems and personalization, the underlying technological and pedagogical mechanisms were not explicitly identified. We argue that clarifying both technological and pedagogical perspectives is important for guiding practical and purposeful SRL scaffolding designs with AI and GenAI.

In this vein, adopting an SRL framework, and particularly, Winne and Hadwin’s (1998) COPES (Conditions, Operations, Products, Evaluations and Standards) model, provides a structured lens for organizing the review. Despite its popularity, to the best of our knowledge, no prior systematic review has adopted COPES as its organizing lens. Collectively, this review aimed to synthesize evidence on the technological deployment of AI and GenAI for SRL scaffolding based on the COPES model, map these interventions onto the SRL process, clarify the distinct and overlapping pedagogical functions, and propose an actionable framework for comprehensive, theory-driven SRL support using generative and hybrid AI.

Literature Review

Winne and Hadwin’s (1998) Model of Self-Regulated Learning

Originating from the information-processing theory, Winne and Hadwin (1998) view SRL as a cognitive–metacognitive process across learning stages. Accordingly, they introduced a recursive four-stage SRL cycle, incorporating the COPES model, and positioned monitoring and control as central components (Figure 1). Particularly, the four SRL stages include task definition, goals and plans, enactment, and adaptation. While other multi-phase cyclical models consolidate initial regulatory processes into a single stage, such as Zimmerman’s (2000) model that positions task analysis and self-motivation beliefs within the forethought phase, or Pintrich’s (2000) framework that combines forethought, planning, and activation in phase one, Winne and Hadwin’s (1998) model separates task definition and goal setting. Greene and Azevedo (2007) contended that this enables the analysis of learners’ initial conditions and intentions. Moreover, the adaptation stage in Winne and Hadwin’s (1998) model involves the learner’s self-reflection on previous stages and long-term changes (Winne, 2018).

Figure 1.

The COPES Model of Metacognitive Monitoring and Control Across Four Stages of Studying. Note. Adapted from Figure 12.1 in “Studying as Self-Regulated Learning,” by P. H. Winne and A. F. Hadwin, in Metacognition in Educational Theory and Practice (pp. 277–304), edited by D. J. Hacker, J. Dunlosky, & A. C. Graesser, 1998, Lawrence Erlbaum Associates

Within Winne and Hadwin’s (1998) model, each of the four SRL stages consistently encompasses the five COPES components of Conditions, Operations, Products, Evaluations, and Standards. Particularly, the learning process starts with Conditions—cognitive and task conditions that guide learners’ cognitive Operations and Standards setting (Winne, 2018). Following this, Operations include learners’ cognitive and metacognitive tactics, with the five basic cognitive SMART operations: searching, monitoring, assembling, rehearsing, and translating (Winne & Hadwin, 1998). These operations generate unique Products at each stage. In parallel, Standards serve as the criteria for learners to judge the quality of these products. Consequently, Evaluations (monitoring and control) compare the products based on standards, informing subsequent adaptations. With the COPES structure, metacognition is positioned centrally in the SRL mechanism as it reflects the continuous evaluation of the products generated and the adjustment of initial cognitive conditions. As such, Panadero (2017) contended that Winne and Hadwin’s framework is the most metacognition-focused of the major SRL models.

Moreover, unlike models that focus primarily on macro-level phases, the COPES architecture enables micro-level analysis of the interactions of components in cognitive architecture across different phases, and presents the recursive dynamics of SRL (Greene & Azevedo, 2007). Moreover, the model explicitly incorporates external evaluations into the SRL mechanism. Taken together, Winne and Hadwin’s (1998) model offers a comprehensive framework for researchers to analyze the participation of AI- and GenAI-based supports in students’ SRL processes and, ultimately, transform them into learning outputs. The model has thus been widely adopted in computer-supported learning research (Panadero et al., 2016).

AI and AI-Based Scaffolding in SRL Development

To articulate the scope of AI, Russell and Norvig (2021) employed the two dimensions of human vs. rational and thought vs. behavior, yielding four approaches: thinking humanely, acting humanely, thinking rationally, and acting rationally. Among the approaches, Russell and Norvig (2021) contended that under the standard model, AI primarily targets rational actions; that is, it chooses the best possible actions to achieve its objectives in a given situation. In the present review, we adopted the rational-agent perspective of AI to identify its scope in supporting SRL, incorporating methods from both human-like behavior (e.g., natural language processing) and rational reasoning (e.g., probability) (Russell & Norvig, 2021). Accordingly, AI includes: (a) knowledge-based systems, (b) uncertainty and probabilistic reasoning, (c) search and optimization, (d) machine learning, (e) natural language processing, (f) computer vision, and (g) robotics (Russell & Norvig, 2021). Moreover, we have further extended this scope to include: (h) multi-agent systems (a subfield of uncertain knowledge and probabilistic reasoning; Russell & Norvig, 2021) to cover scenarios involving multiple agents pursuing individual and collective goals.

This section presents the main ideas and representative keywords of the eight AI subfields based on the framework of Russell and Norvig (2021). One major area is machine learning (ML), which refers to systems learning from data to improve their performance. It includes supervised, unsupervised, and reinforcement learning, with deep learning widely used across these approaches. Another area is knowledge-based systems, which rely on formal rules and logic to represent knowledge and support decision-making. Common terms associated with this area include logical agents, knowledge representation, automated planning, rule-based systems, and expert systems. Uncertainty and probabilistic reasoning focus on how AI systems deal with incomplete or uncertain information when making decisions. This area includes models such as basic probability models, Bayesian networks, Hidden Markov Models, and fuzzy logic. Natural language processing (NLP) concerns how systems understand and generate human language. Its common tasks include speech recognition, text-to-speech, machine translation, information extraction and retrieval, and question answering. Search and optimization involve algorithms that examine different possible options and identify the most suitable solution. Common approaches include state-space search, optimization-based methods, constraint satisfaction, and adversarial search, such as in games. Robotics refers to machines that can sense, reason, and act in the physical world. Computer vision focuses on helping AI interpret visual information, such as images and objects. Finally, multi-agent systems assign different tasks to multiple agents with core tasks ranging from planning to coordination. Table 1 summarizes the concepts and keywords of these eight subfields.

Table 1.

AI Subfields

AI subfields	Concepts	Typical keywords
Knowledge-Based Systems	Systems to store and manipulate information of logic and ontologies	Logic, knowledge representation, knowledge tracing, automated planning, rule-based system, and expert systems
Search & Optimization	Algorithms to find solutions or make decisions by exploring possible options	State-space search, optimization-oriented representations, constraint satisfaction problems, and adversarial search (games)
Uncertainty & Probabilistic Reasoning	Probability theory to handle uncertainty in knowledge and decision-making	Bayesian network, Markov models, and fuzzy logic
Machine Learning (ML)	Methods that allow computers to learn from data and improve over time	Supervised, unsupervised, reinforcement, and deep learning
Natural Language Processing (NLP)	Techniques for understanding and generating human language	Speech recognition, text-to-speech, lexical, machine translation, information extraction and retrieval, and question answering
Robotics	Machine with perception, reasoning, and action to physically operate	Perception, planning, and motion control
Computer Vision	Machine to interpret and understand visual information from images or videos	Image processing and object recognition
Multi-agent System	Systems where multiple intelligent agents interact or collaborate	Planning and cooperation

Note. Adapted from Artificial Intelligence: A Modern Approach (4th ed.) by S. Russell and Norvig (2021), Pearson.

To date, diverse AI technologies have been used in SRL development. Particularly, ML has been employed for final performance predictions (Guerrero-Roldán et al., 2021). Intelligent tutoring systems (ITSs), such as MetaTutor (Duffy & Azevedo, 2015), provide data-driven adaptive scaffolding across the SRL process. Other systems, such as DSLab-Bot (Ortega-Ochoa et al., 2024), combine multiple AI technologies to detect students’ emotions, and hence can support both cognitive and emotional aspects of SRL.

GenAI and GenAI-Based Scaffolding in SRL Development

GenAI, by definition, is a subfield of deep learning and was designed to generate new content by learning from existing datasets (Tomczak, 2022). Creative intelligence distinguishes GenAI from other AI technologies, particularly discriminative AI, which focuses solely on analyzing and classifying existing data to make predictions or support decision-making (Banh & Strobel, 2023). GenAI applications encompass various modalities of text, audio, video, code, image, and others (Banh & Strobel, 2023) with large language models (LLMs) centering recent advancements (Yang et al., 2025). Regarding architecture, GenAI can be accessed through: (a) direct access to standalone websites or apps, (b) open-source or commercial models via an application programming interface (API), or (c) fully integrated end-to-end systems (Banh & Strobel, 2023).

In terms of model customization, the base model can be further customized by: (a) fine-tuning that refines the base model using additional datasets, (b) in-context learning that refers to users and system developers providing specific prompts or examples without altering the model’s underlying parameters, and (c) input and output filtering that moderates or blocks certain prompts or responses (Ohm, 2024). Of those, prompt engineering is a well-known form of in-context learning. It involves designing and iteratively refining prompts to elicit desired model output (Ohm, 2024).

Moreover, LLM-based GenAI can also be integrated with external knowledge bases. In particular, basic knowledge bases (KBs) provide structured repositories for predefined facts, relationships, and rules; knowledge graphs (KGs) enable interconnected entities and relationships; and retrieval-augmented generation (RAG) provides real-time access to up-to-date information and dynamic content (Yang et al., 2025). Table 2 summarizes the various GenAI configurations.

Table 2.

GenAI Configurations

GenAI	Characteristics	Concept	References
Modality	Text	X-to-text	Banh and Strobel (2023)
	Audio	X-to-audio
	Video	X-to-video
	Image	X-to-image
	Code	X-to-code
	Other	X-to-molecules models
Architecture	Direct Access	Direct interaction with GenAI through standalone websites or applications	Banh and Strobel (2023)
	Open applications	Utilize open-source or commercial models via APIs (e.g., Llama 3.1)
	End-to-end application	Fully integrated systems with full control, in specialized domains (e.g., GitHub Copilot)
Model customization	Fine-tuning	Refine the base model using additional datasets	Ohm (2024)
	In-context learning	Users/system developers provide specific prompts or examples without altering the model’s underlying parameters A well-known form: Prompt engineering
	Input and output filtering	Moderates or blocks certain prompts or responses
External knowledge bases	Basic knowledge bases (KBs)	Structured repositories storing predefined facts, relationships, and rules	Yang et al. (2025)
	Knowledge graphs (KGs)	Interconnected entities and relationships
	Retrieval-augmented generation (RAG)	Real-time access to up-to-date information and dynamic content

GenAI has increasingly been leveraged to scaffold SRL. For example, Pan et al. (2025) integrated ChatGPT via API with prompt engineering to generate personalized reading content and elaborate self-regulated reading strategies. Another chatbot (XtagGPT via API) sends assignment reminders, asks about students’ weekly progress, recommends help resources, and guides strategic planning (Tsai et al., 2025).

Previous Reviews and Related Work

Existing systematic reviews in AI–SRL and GenAI–SRL research have primarily focused on intervention outcomes rather than design characteristics. For AI–SRL in language learning, Chang and Sun (2024) showed that AI interventions effectively support goal setting, self-monitoring, self-reflection, and strategy use; however, most interventions have targeted a single phase of SRL. Sardi et al. (2025) reported that most GenAI interventions have positive impacts on students’ SRL, primarily through adaptive feedback, personalization, and metacognitive support. However, concerns remain about overreliance and possible harm to independent thinking. In this sense, purposeful and flexible SRL scaffolding designs are necessary.

Several conceptual papers have proposed principles for intentional design of GenAI for SRL. For example, Chang et al. (2023) suggested three key features for designing GenAI chatbots to support SRL: goal setting and prompting, self-assessment and feedback, and personalization and adaptation. They also noted that designers should pay attention to pedagogical aims, ethical issues, and academic integrity. Although these ideas offer a useful starting point, they are mainly based on conceptual reasoning and a limited set of prior studies. More systematic reviews with formal evidence synthesis are therefore still needed.

As the only study to date that integrates GenAI into the broader AI stream when examining the AI–SRL relationship, Banihashem et al.’s (2025) review provided an overview of participants, theory, implementations, and applications. The reported implementations mainly involved adaptive systems and personalization, prediction and profiling, intelligent tutoring systems, and assessment and evaluation. However, their review did not explicitly report the underlying technologies or the specific pedagogical purposes of the SRL process. Additionally, they placed insufficient emphasis on GenAI interventions despite their recent widespread employment.

Adopting a process view of SRL as learner-initiated regulation, our study aimed to: (a) synthesize evidence on AI- and GenAI-based support for SRL from both technical and pedagogical perspectives, and (b) propose an actionable framework for more comprehensive and theory-driven SRL scaffolding with AI and GenAI. We adopted Winne and Hadwin’s (1998) model of SRL as the organizing lens for this review.

The current study addressed the following research questions:

(1) What AI technologies and GenAI configurations have been used for scaffolding SRL?

(2) How convergently and divergently have GenAI and AI scaffolded the SRL process based on the COPES model?

(3) How can GenAI and AI more comprehensively support the SRL process?

Methods

This systematic review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (PRISMA; Page et al., 2021; Figure 2). The procedure included the stages of identification, eligibility criteria, screening process, and data analysis and categorization. We also conducted a quality assessment of the included articles in the systematic review.

Figure 2.

Selection process based on PRISMA

Identification

Database searches were conducted in the Web of Science, Scopus, and ERIC databases to identify articles included in the systematic review. All searches were conducted in May 2025. Moreover, the reference lists of recent reviews were also checked for additional articles. Regarding the search string, the study used three keyword categories: AI, GenAI, and SRL. The Boolean OR linked the search terms for each category. A full search string for AI-based SRL scaffolding search included keywords of AI and SRL topics, connected by the Boolean AND; a similar string was used for the GenAI articles. Appendix 4 provides detailed searching strategies for each database platform.

Eligibility Criteria

The eligibility criteria were established to identify articles included in the review (Table 3). We selected the timeframe from 2015 to 2025 based on the results of a previous review, which showed that AI-in-education publications in the Web of Science took off around 2015 and continued to accelerate after 2018–2019 and from 2022 onward (Durak et al., 2024). Moreover, we included empirical studies that implemented AI- and/or GenAI-based SRL scaffolding interventions and reported an evaluation (quantitative, qualitative, or mixed-methods). This was to ensure empirical evidence of purposeful design choices, perceived usefulness, and implementation.

Table 3.

Inclusion and Exclusion Criteria

No	General criteria	Inclusion	Exclusion
1	Published Year	Between 2015 and 2025	Before 2015
2	Published Language	English	Other than English
3	Publication Type	Peer-reviewed studies	Non-peer-reviewed sources
4	Study Design	Empirical intervention studies reporting evaluation data (quantitative, qualitative, or mixed-methods)	Non-empirical or non-evaluation papers (e.g., conceptual papers, reviews)
5	AI- or/and GenAI-based Design for SRL	Specific AI- or/and GenAI-based design to scaffold SRL	Nonspecific SRL design
			Non-AI or non-GenAI
			Duplicated systems
6	Paper Accessibility	Full-access papers	Not fully accessible

Screening Process

The screening procedure underwent the following steps:

(a) Consolidating all identified articles into a single Excel file,

(b) Removing duplicates manually by comparing author names, titles, and publication details, and

The initial database contained 3,419 articles, which were reduced to 2,291 after duplicates were removed. Thereafter, titles and abstracts were screened against the inclusion and exclusion criteria, resulting in 180 articles. Moreover, from the reference lists of recent reviews, we identified two additional articles, bringing the total to 182 for full-text assessment. Finally, 70 articles (representing 72 studies) were included in the systematic review, with 38 AI-based and 34 GenAI-based studies (two articles counted for both, given their hybrid designs) (Figure 2). Included studies comprised 56 studies with experimental/quasi-experimental evaluations and 14 non-experimental qualitative, quantitative, or mixed-method evaluations.

Quality Assessment

The study adopted the Mixed Methods Appraisal Tool (MMAT; Hong et al., 2018) to assess the quality of the articles included in the systematic review. In the first round, one author conducted the quality assessment of each article, following predefined and transparent criteria and instructions for equivalent methodology to rate each criterion as “met,” “not met,” or “cannot tell.” The overall assessment for each article was then categorized as low (<50% criteria met), moderate (50–75%), or high (>75%) (Asfani & Chen, 2025). Subsequently, a second author randomly selected and independently reviewed 15 (21%) of the included articles to verify the initial quality appraisal. Agreement on the overall quality category was 100% for the verified subset; all 70 articles rated as moderate to high quality (Table 4), and thus all were considered suitable for inclusion in the review.

Table 4.

MMAT Quality Assessment

MMAT quality assessment rating	Score range (%)	Number of articles	Example article
High	>75	39	Afzaal et al. (2024)
Moderate	50–75	31	Matsuda et al. (2020)

Data Categorization and Analysis Using Coding Schemes

This review adopted the deductive and inductive content analysis methodology (Elo & Kyngäs, 2008). The coding process applied frameworks derived from AI and GenAI typologies and the COPES model. Extracted qualitative data included information on the research context, domain, technology, and system design. The analysis process was conducted using Microsoft Excel.

AI and GenAI Technologies Extraction and Categorization

Data Extraction

We scrutinized the research manuscripts and supplementary documents to extract the authors’ reported technical descriptors or model names. For example, the AI technology used in the research by Han et al. (2026) was recorded as “Dialogflow with a rule-based framework.” For the GenAI group, we extracted technical specifications according to four GenAI configurations (Table 2). For example, Wu et al. (2024) reported using “GPT-3.5, via Apple Shortcuts and LINE integration,” which we coded as the system’s modality and architecture. When the included study did not specify the underlying technology, the researchers sought related publications and publicly available materials (e.g., project pages or the tool’s websites) to obtain additional technical details. Moreover, if the technical evidence was insufficient in both the article and external materials, we coded the technology as “unclear.” For instance, Mosoteach used in Liu et al.’s (2023) study integrates AI to analyze behavioral data and provide personalized teaching and learning suggestions. However, the specific AI technology underlying the tool was not reported in either the article or the tool’s official website. Therefore, the technology was coded as “unclear.” Additionally, for the GenAI group, if a study did not utilize model customization or knowledge bases, the corresponding fields were left blank in the extracted dataset. Data extraction is recorded under the column “Reported Technology” in Appendix 1.

Data Categorization

For the AI group, we mapped each study’s reported technical description to one or more of the eight predefined AI technology categories (Table 1). For instance, in Han et al.’s (2026) study, the reported technology, “Dialogflow, rule-based framework,” was classified as “ML, NLP, and knowledge-based systems.” For the GenAI group, we extracted data by GenAI configurations and consolidated them into a single entry to represent the system’s overall technical profile. For example, in Wu et al.’s (2024) paper, modality was recorded as “text,” architecture as “via API,” model customization as “no model customization,” and external knowledge bases as “no KB”; therefore, the mapped technology was “Text; via API; no model customization; no KB.” If a study employed multiple technologies, these were recorded within the corresponding cell. Data categorization is recorded under the column “Mapped Technology” in Appendix 1.

COPES Coding Schemes

We adopted the COPES model (Winne & Hadwin, 1998) to analyze the pedagogical purposes. We scrutinized the research manuscripts and supplementary documents to extract content of AI- and GenAI-based SRL scaffolding. Subsequently, we employed deductive content analysis to categorize the extracted content by SRL stages (task definition, goals and plans, enactment, and adaptation) by the specific SRL process the intervention was intended to support, and by COPES components (conditions, operations, products, evaluations, and standards) for the primary instructional purpose of the intervention. For example, a GenAI system by H. Li (2023) offered multiple supports, including automatic daily learning tasks and task reminders, delivered learning content in response to students’ inquiries, sent reminders that encouraged users to engage in critical conversations and to be cautious about potentially incorrect ChatGPT-generated feedback, while also summarizing students’ learning participation (e.g., number of posts). In this case, we coded the daily task announcement and reminders as learning management supports because they support students’ scheduling of learning activities, but the reminders that prompt critical discussion and caution about AI-generated feedback serve as metacognitive scaffolding, as they encourage students’ self-monitoring and evaluation during the task enactment. The automated summary of students’ learning participation (e.g., number of posts), however, was coded under the adaptation stage because it was intended to support later adjustments rather than immediate task performance. Lastly, support for learning content in response to students’ inquiries was coded as Q&A and learning content generation, serving as a source of learning content and support for help-seeking. Moreover, in an AI system (Long & Aleven, 2017), a prompt asking students to rate their confidence in their answers to equation-solving questions before submission was coded as metacognitive scaffolding because it supported student self-monitoring rather than providing evaluative information. Accordingly, this was classified under enactment within operations. In contrast, the real-time skill bars that displayed the estimated mastery progress were coded as metacognitive feedback during enactment within evaluations. Subsequently, multiple supports, within the same category (e.g., cognitive scaffolding), were counted once if they shared the same primary instructional purpose. However, the same intervention, when coded across different phases, was counted separately for each phase. For example, in Mills et al. (2025), cognitive scaffolding for goals and plans (teaching on weekly tasks and SRL selection) was counted once, and cognitive scaffolding during enactment (recommending SRL activities based on online behavior) was counted as a separate instance.

To ensure the reliability of COPES’s coding, two authors independently coded 21% (n = 15) of the 70 included papers, covering as much content as possible. Discussion sessions between authors, followed by individual self-review, continued until full consensus for the predefined subcodes was achieved. Intercoder reliability was high (Cohen’s κ = .80). Thereafter, the first author coded the remaining articles, with discussion of any newly emerging subcodes.

The following part presents the COPES coding schemes from the reviewed studies. All coded interventions and their COPES mapping are provided in Appendix 3. The Conditions facet captured external sources from which subsequent support could be generated. All interventions reflected aspects of Conditions, as technologies established the learning environment and provided sources for subsequent operations, products, evaluations, and standards. Specifically, we coded: (1) source of knowledge for providing learning tactics, task knowledge, or subject-matter knowledge available to learners, and (2) source of instructions for instructional guidance during task performance.

The Operations facet encompassed scaffolding that directly supported learners’ task engagement. Within Operations, we coded: (3) cognitive scaffolding for interventions that teach the task and system use, domain knowledge, and SRL skills, and help learners organize and elaborate learning content, (4) metacognitive scaffolding for prompting students to set goals and monitor understanding, evaluate progress, and adjust strategies, and (5) affective scaffolding for enhancing learners’ awareness and regulation of affective factors. We coded (6) Q&A and learning content generation for system-generated learning content or responses to learner queries, highlighting their role as learning resources and help-seeking supports. Moreover, we coded (7) test and exercise generation for the generation of tests and exercises on ongoing learning content and transfer tasks. Notably, we treated (8) automated metacognitive reports as distinct from metacognitive scaffolding, as they involve reflective analytics which are most closely aligned with end-of-cycle evaluation supporting adaptation, rather than in-task scaffolding. Moreover, there were (9) learning management supports for scheduling, organizing, and tracking learning activities, (10) negotiated help-seeking regulation, and (11) teachable agent interactions.

The Evaluations facet consisted of system-generated information that appraises performance or learning states and supports monitoring and control. We coded (12) grading as involving automatically scoring students’ work, (13) cognitive feedback for the task-level evaluation of subject matter, (14) metacognitive feedback for the meta-level evaluation supporting self-reflection, monitoring, and regulation, and (15) affective feedback for feedback aimed at reducing learners’ anxiety and maintaining engagement. The Products facet captured the tangible outputs generated by the system or co-constructed through learners’ engagement. We coded (16) learning path products for the generation and updating of goal-based sequences of learning steps and activities, and (17) assignment production generation for the generation and refinement of learner work products. Lastly, the Standards facet included explicit criteria for judging performance and guiding regulation. We coded (18) benchmark setting for defining required performance thresholds and (19) task prioritization for defining priorities or sequencing to achieve the task.

Finally, we employed Python and Microsoft PowerPoint to visualize the study’s findings.

Results and Discussion

Background Information

This section provides an overview of the 70 reviewed publications. Our analysis revealed a growing scholarly focus on AI- and GenAI-based SRL, with publications between 2024 and 2025 accounting for 70% (n = 49). This finding shows how GenAI–SRL research has rapidly expanded since ChatGPT’s release, in parallel with the continued development of earlier AI systems. Moreover, consistent with Banihashem et al. (2025), we found that most research was conducted in higher education (n = 52), followed by K–12 (n = 16) and lifelong learning (n = 2). In terms of domain, work mainly focused on Language Skills (22.5%), Electrical Engineering & Computer Science (21.1%), STEM Education (18.3%), Health & Nursing (12.7%), and Education (8.5%) (Appendix 1).

Against this backdrop of rapid growth and diverse application contexts, the present review goes beyond aggregating descriptive features by offering a comparative, process-oriented explanation of how earlier AI and GenAI instantiate different regulatory mechanisms across the SRL process. Specifically, we mapped technological foundations and pedagogical applications onto Winne and Hadwin’s COPES architecture and SRL stages, enabling us to: (a) locate where supports concentrate and where they remain sparse, (b) articulate distinct affordances of analytics-driven AI versus dialogic, context-adaptive GenAI, and (c) identify systematic gaps in current AI- and GenAI-based designs for SRL scaffolding. These comparative insights are synthesized into the proposed framework presented in RQ3.

RQ1. What AI Technologies and GenAI Configurations Have Been Used to Scaffold SRL?

Among the included publications, 34 of 38 AI articles and 33 of 34 GenAI articles reported technical details. Figure 3 shows an overview of AI and GenAI technologies used to support SRL, grouped by publication year. Overall, GenAI-based SRL systems grew rapidly in 2024 and 2025. There were nine studies in 2024, which rose sharply to 25 in 2025, suggesting increasing interest in using GenAI to support SRL systems. In contrast, earlier AI approaches appeared over a longer period of time. Particularly, knowledge-based systems were used fairly consistently from 2017 to 2025, and machine learning and natural language processing gained popularity in 2024 and 2025.

Figure 3.

Trends in AI-based and GenAI-based SRL support by publication years. Note. The total count of mapped technologies exceeds the number of articles because many systems integrate multiple technologies

Moreover, computer vision, search and optimization, and uncertainty and probability reasoning appeared less frequently and were primarily reported in more recent years. Overall, the trends suggest that AI has continued to be utilized alongside GenAI rather than being replaced, with growing scholarly attention to multiple dimensions of SRL support (e.g., affective scaffolding and activity recommendations). The following part describes in detail AI technologies and GenAI configurations and the SRL scaffolding output they enable. Full citations for the literature on AI technologies and GenAI configurations are provided in Appendix 2.

AI Technologies for Scaffolding SRL

Knowledge-based systems, machine learning, and natural language processing were the main AI technologies used for SRL scaffolding. Other technologies included uncertainty and probabilistic reasoning, search and optimization, multi-agent systems, and computer vision (Figure 3). Full literature citations by AI technologies are provided in Appendix 2.

Knowledge-based systems (n = 16) were widely used for their logic transparency, aligning system decisions with predefined instructional rubrics and pedagogical principles. In particular, most studies implemented rule- or script-based reasoning to deliver structured guidance, adaptive feedback, and real-time, personalized, and principle-driven next steps based on learners’ evolving responses, confidence, and behavior (e.g., Al-Shanfari et al., 2020; Du et al., 2025; H. Liu et al., 2024). A smaller subset adopted an expert-novice overlay model to compare the learner’s status with the expert’s complete set of knowledge components and to visualize skill meters (T. Wang et al., 2023). Collectively, these systems enabled SRL scaffolding by supporting goal setting and planning, ongoing, timely, diagnostic, and procedural guidance, feedback, help-seeking regulation, and progress visualizations and recommendations for monitoring and adaptation.

ML (n = 15) was largely utilized to enhance SRL systems’ adaptivity by learning patterns from learner traces (e.g., assignment scores, text, and speech) and refining system guidance over time. Specifically, several studies employed ML to build predictive models of learning outcomes, together with strategy recommendations to trigger students’ self-monitoring and self-reflection (Afzaal et al., 2024; Guerrero-Roldán et al., 2021). Other work used ML to compute performance thresholds for adaptive instructional specificity or challenge in game-based learning (Ferguson et al., 2022) or to enable learning-by-teaching designs via teachable agents (Matsuda et al., 2020). Notably, ML has been frequently integrated with NLP to enhance adaptive, accurate analysis. For example, ML was used in written and spoken language, and in image and knowledge graph analysis, to deliver real-time feedback and adaptive content and difficulty levels (Shafiee Rad, 2025). In other ML-powered NLP systems, ML supported sentence classification, error detection, feedback generation in language writing tasks (e.g., Neshaei et al., 2025), retrieval-based response selection (Ortega-Ochoa et al., 2024), and cognitive diagnosis (Liao et al., 2024). Taken together, ML supported SRL scaffolding by offering predictive performance models, performance thresholds, pedagogical agents, and improving the accuracy of analysis and feedback.

NLP (n = 15) enabled systems to understand, interpret, and generate human language. Most commonly, ML-powered NLP, as discussed above, permitted the classification of learners’ input and the generation of adaptive learning content and feedback. Previous studies also utilized standalone NLP or NLP integrated with other systems. For example, an NLP-based system handled online test and review session booking and student request tickets regarding course content and materials based on the appropriate answers available in the database (Al-Abdullatif et al., 2023). Moreover, Perikos et al. (2017) combined NLP with a knowledge-based system to deliver different automatically generated natural language hints and feedback sequences that were adapted to students’ metacognitive state before and after assignment submissions. In general, NLP served as a core linguistic interaction component in SRL systems, translating learners’ free-text inputs into actionable signals, and delivering adaptive and personalized learning content, instructions and prompts, feedback, learning administration, and recommendations.

Uncertainty and probabilistic reasoning (n = 4) enabled SRL systems to decide learners’ current states (e.g., mastery, uncertainty, engagement, and affect) from observable traces, thereby driving adaptive instructions, feedback, and progress updates. For example, one system combined fuzzy logic with NLP, ML, and a knowledge-based system to provide content-related, adaptive cognitive and affective feedback (Ortega-Ochoa et al., 2024). However, in a standalone uncertainty and probabilistic reasoning system, Long and Aleven (2017) leveraged Bayesian Knowledge Tracing with Knowledge Components to enable students to select the subsequent problem level, view overall and skill progress via dashboards, and receive self-assessment prompts and delayed feedback.

Moreover, search and optimization (n = 2) primarily supported SRL by adapting task sequencing and learning content to learners’ goals, self-efficacy, and ongoing progress. For example, one system selected instruction and content based on students’ self-assessment and performance reports (Vertemati et al., 2024). Another ML-based search and optimization system dynamically sequenced learning tasks around learners’ goals, and offered real-time recommendations (Afzaal et al., 2024). Multi-agent systems (n = 1) allowed different SRL subprocesses to be externalized and scaffolded through specific pedagogical agents, such as sub-goal prompts, self-evaluation questions, and strategy reminders (MetaTutor; Duffy & Azevedo, 2015). Computer vision (n = 1) extended SRL support to the affective dimension by detecting learners’ affect and attention from facial and emotional cues. These cues were subsequently analyzed using ML to identify affective states and provide adaptive affective feedback (i.e., encouragement, praise, or alternative tasks) to reduce anxiety and sustain engagement (H.-L. Liu et al., 2022).

GenAI Configurations for Scaffolding SRL

GenAI-for-SRL models were configured through different combinations of modalities, architectures, customization, and external knowledge bases (Figure 4(a)). Except for one study with unspecified models, all other reviewed GenAI studies (n = 33) utilized LLMs for text generation, with ChatGPT being the most widely used, followed by Llama 3.1 and BaiChuan. Text-based modalities primarily produced dialogic outputs, functioning as learning content, instructions, feedback, prompts, and hints (e.g., Mzwri & Turcsányi-Szabo, 2025). Text modalities were also embedded in the backend analysis systems (e.g., Jin et al., 2025). Additionally, audio-to-text and text-to-speech models (n = 2) (i.e., OpenAI Whisper and ElevenLabs) were used in Mills et al.’s (2025) study for speech transcription and synthesis so that the AI Companion could provide voice-based guidance throughout VR-supported French writing activities, varying from model analysis to scaffolded feedback. In general, the dominance of text-based LLMs suggests that GenAI-based SRL systems mainly operate through dialogic and language-mediated processes rather than through embodied or sensory-rich regulation.

Figure 4.

GenAI Configurations and Co-Occurrences in SRL Scaffolding: (a) individual GenAI configurations; (b) co-occurences of GenAI configurations. Note. One study with unclear technology was excluded from the figure, and Mills et al. (2025) utilized three modalities: text generation, speech transcription, and speech synthesis

In terms of system architecture, most GenAI systems were connected through APIs (n = 28), suggesting that researchers more often embedded GenAI into structured learning environments where support could be better aligned with the task and learner profiles (e.g., M. Liu & Reinders, 2025). Meanwhile, a smaller number of studies (n = 7) used GenAI through standalone apps or websites. In these cases, the tools were mainly used when learners chose to seek support on their own, but they were generally part of a less systematic instructional design (e.g., S. Wang et al., 2024).

As for model customization, in-context learning, particularly prompt engineering (n = 26), was a widely adopted technique that permits GenAI’s output to align with learning goals, task requirements, and learners’ regulatory needs. Prompt engineering could be implemented through two methods: (a) at the system level without altering the model’s underlying parameters (e.g., Nguyen et al., 2025), and (b) students’ direct use and modification of provided prompt examples (e.g., Sun et al., 2025). In contrast, few studies (n = 2) used fine-tuned methods to train the pre-trained model with domain-specific SRL and course content (e.g., Gao et al., 2025). The wide adoption of prompt-engineering over fine-tuned methods reflects a priority for flexible and rapid deployment for surface-level alignment over deep domain adaptation in existing GenAI-based SRL systems.

Knowledge bases were not integrated in most current systems (n = 25; e.g., Mzwri & Turcsányi-Szabo, 2025). Among the studies that included knowledge bases (n = 8), most used basic forms, such as user databases, task modules, or content modules. They were mainly used to provide personalized support, monitor participation, give immediate feedback, and send reminders or assign tasks (e.g., Hu et al., 2024; H. Li, 2023). Only two studies used retrieval-augmented generation (RAG) to provide real-time, discipline-specific content (e.g., Neumann et al., 2025). Overall, the limited external knowledge bases in existing GenAI-SRL systems suggest that they still functioned more as conversational agents than as learner-centered systems grounded in subject knowledge.

Current GenAI systems supported SRL through different combinations of the above components. In some studies (n = 6), GenAI was used via direct access for text generation, with prompt engineering, without linking to a knowledge base (e.g., Campos, 2025). In these cases, simple prompt designs, such as templates, role prompts, step prompts, and rubrics, played a central role in shaping both the generated content and the SRL scaffolding. However, more systems were built through API integration, with prompt engineering but without knowledge bases (n = 12). In these systems, GenAI worked at the backend, and prompts, prominently system-initiated prompts, translated pedagogical intentions into a controlled, predefined scaffolding sequence. Some other API-based GenAI systems (n = 4), however, relied on basic knowledge bases without model customization. In these systems, knowledge sources were mainly used to support learners’ cognitive and metacognitive processes or to provide personalized guidance and assignments (e.g., Hu et al., 2024; H. Li, 2023). In addition, a few API-based systems (n = 6) combined both prompt engineering and knowledge bases. With this design, prompt engineering ensured the generation of the right kind of SRL scaffolding, for example, the four-level scaffolding outputs (e.g., A. Y. Q. Huang et al., 2025) or targeted metalinguistic guidance and timely feedback for reflection (M.-R. Chen, 2024). The integration of learning material and task databases further allows prompts and responses to be anchored to the course workflow (e.g., materials, assignments) and can reduce the risk of hallucination. The co-occurrences of GenAI configurations are illustrated in Figure 4(b).

Summary

Grounded in rule-based algorithms, learning analytics, natural language processing, and computer vision, AI technologies analyzed a wide range of inputs (e.g., students’ inputs, real-time log data, and behavioral traces) in their analytical intelligence in combination with pedagogical rules to generate diverse data-driven SRL support, aligned with students’ learning processes. AI-based SRL systems mainly delivered structured guidance, adaptive hints, feedback, and strategy recommendations; goal-setting and planning prompts; self-monitoring and evaluation support; progress tracking; benchmark setting; performance prediction; learning administration support; pedagogical agent interactions; and responses to students’ help-seeking requests and personalized content generation. However, system outputs were occasionally perceived as lacking contextual awareness, flexibility, and interactive features, and as being less helpful for learners who already had strong SRL skills (Han et al., 2026; Hew et al., 2023).

In the GenAI group, most studies used ChatGPT, often via API integration and prompt engineering to better match learners’ goals and self-regulatory needs. Only a small number of systems used external knowledge bases to improve accuracy and make the support more context-specific. In general, GenAI-based SRL support in the included studies was predominantly text-based and conversational; it excels at delivering human-like, personalized, and context-aware outputs for instructions, prompts, feedback, learning content, visualizations, and learning administration. Nevertheless, GenAI-based chatbots were occasionally reported to be passive, with mismatched feedback, and to lack autonomous reminders (Ng et al., 2024).

In response to Research Question 1, we have described each AI technology and GenAI configuration and the SRL scaffolding output they enable. Moving beyond technical descriptions, to respond to Research Question 2, we systematically examined the pedagogical purposes of these SRL interventions, using Winne and Hadwin’s (1998) model.

RQ2. How Convergently and Divergently Have AI and GenAI Scaffolded the SRL Process Based on the COPES Model?

Building on the deductive and inductive content analyses of AI- and GenAI-based interventions, this section examines where and for what pedagogical purposes these interventions operate within the SRL process. Using Winne and Hadwin’s COPES as the organizing lens for this systematic review, we mapped the coded interventions to: (a) COPES facets (Conditions, Operations, Products, Evaluations, and Standards) and (b) SRL stages (task definition, goals and plans, enactment, and adaptation). We also analyzed how AI and GenAI scaffolded the SRL process convergently and divergently. In general, interventions across both technology groups were most concentrated in Operations and Evaluations. They were predominantly situated in the enactment stage, with comparatively limited attention to task definition, goals and plans, and adaptation. Full literature citations for each intervention, organized by COPES facets and SRL phases, are provided in Appendix 3.

AI- and GenAI-Based SRL Scaffolding in Conditions

Both AI and GenAI acted as sources of knowledge and sources of instructions. However, GenAI in the included studies focused more on knowledge provision (n = 44; e.g., content and test generation, conversational elaboration on informational requests), whereas AI more often supported instructions (n = 47; e.g., metacognitive prompts based on students’ progress).

AI- and GenAI-Based SRL Scaffolding in Operations

Cognitive scaffolding was prevalent across both AI- and GenAI-based systems, occurring at comparable frequencies (AI: n = 13; GenAI: n = 17) and prominently during task enactment. During task enactment, GenAI-based cognitive scaffolding was often co-constructed through conversation and tailored to learners’ immediate responses; thereby, the scaffolds were typically dynamic, dialogic, learner-driven, and context-sensitive. Existing GenAI-based cognitive scaffolding during enactment included problem and task analysis, misconception diagnosis, revision guidance, writing and editing assistance, follow-up questions and instructions, query suggestions for interacting with the system, knowledge clarification, provision of external resources, and learning recaps (e.g., Jin et al., 2025; Lai et al., 2025; Lee et al., 2024; Pan et al., 2025; S. Wang et al., 2024). A smaller subset (e.g., A. Y. Q. Huang et al., 2025) facilitated four levels of scaffolding, from flowcharts to advanced solutions for a programming learning system, or visualized knowledge in the form of knowledge maps and knowledge cards (Gao et al., 2025). By contrast, AI-based cognitive scaffolding mainly took the form of structured progression control, including predefined pop-ups, hint levels, next-step recommendations, prompts, adaptive difficulty adjustments, mastery learning pacing, next-step unlock, and teaching learning content with Q&A (e.g., Arvidsson & Kuhn, 2021; Du et al., 2025; McCarthy et al., 2018; Ng et al., 2024). In the other SRL phases, cognitive scaffolding mainly helped learners with task analysis and system use, prior knowledge review, and learn to set goals and plans (e.g., GenAI: Hu et al., 2024; AI: Du et al., 2025; Glick et al., 2024). In general, the differences in how the two groups scaffolded cognition align with their core mechanisms. GenAI systems were better able to support dynamic knowledge construction, while AI systems relied on predefined algorithms and databases to decide when to support and select content to offer.

Metacognitive scaffolding in both technology groups covered the goals and plans, enactment, and adaptation phases, with the most support focused on task enactment. The AI group appeared more often than the GenAI group in this regard (AI: n = 24; GenAI: n = 11). Specifically, AI-based metacognitive scaffolding appeared more diverse and evenly distributed across the four SRL phases. Typical interventions included prompts for structured goals and plan setting, deployment of learning strategies and self-monitoring, hint panels for next steps, opportunities for self-regulatory selections of questions or activity levels, satisfaction and confidence ratings, structured self-reflection guidance, follow-up activity recommendations, and visualizations of task progress updates (e.g., Duffy & Azevedo, 2015; Lim et al., 2024; Long & Aleven, 2017; Vertemati et al., 2024; Yin et al., 2025). In contrast, GenAI-based metacognitive scaffolding was more dialogic and dynamic in response to learners’ immediate inputs, serving as sustained coaching dialogue that guided goal setting, strategy selection, iterative strategy reevaluation and resource seeking, self-reflection and planning for improvement, and notably critical metacognition regarding risk of AI-generated hallucinations (e.g., Lin et al., 2025; M. Liu & Reinders, 2025; Sun et al., 2025; M. Zhu, 2025). Consistent with cognitive scaffolding patterns, AI-based metacognitive scaffolding was typically analytics-triggered and appeared more diverse and closely aligned with learners’ behavioral data, whereas GenAI-based scaffolding primarily comprised conversation-based scaffolds following learners’ inputs.

Affective scaffolding has remained limited in both technologies, and was implemented during the enactment and adaptation stages (AI n = 3; GenAI n = 1). A GenAI system included motivational quotes and basic emotional support (S. Wang et al., 2024), and an AI system adopted emojis and shared reflections from previous learners to foster social presence (e.g., Hew et al., 2023). The scarcity of affective scaffolds in both technologies suggests a design gap in supporting learners’ affective awareness and regulation throughout the SRL process.

Q&A and Learning Content Generation highlights the role of just-in-time content and resources in supporting learners during the learning process, and was concentrated during task enactment in both technology groups. However, this function in the reviewed studies was much more prevalent in GenAI than in AI (AI: n = 3; GenAI: n = 19), reflecting GenAI’s stronger capacity for open-ended, context-sensitive generation. GenAI systems typically generated reading passages, learning content, and Q&A responses based on learners’ prompts and interaction contexts, sometimes incorporating learner-specific filters (e.g., language proficiency, output complexity), and increasingly grounding outputs in external knowledge sources, such as knowledge bases, uploaded materials, or other real-time resources, to improve learner- and discipline-specific accuracy (e.g., Neumann et al., 2025; Pan et al., 2025; Xu et al., 2024). In contrast, AI-based implementations in this category were less common and tended to rely on selecting or delivering pre-prepared content in response to learners’ queries or assigning suitable learning materials based on learners’ self-competency ratings (e.g., Ortega-Ochoa et al., 2024; Vertemati et al., 2024). Overall, the stronger presence of GenAI-based Q&A and learning content generation support suggests a shift toward GenAI as a more flexible and responsive resource for learning content and help-seeking.

Test and exercise generation occurred primarily during enactment and adaptation, with comparable frequencies between GenAI and AI (AI: n = 5; GenAI: n = 7). Reviewed GenAI-based systems tended to generate new, just-in-time items, such as structured passage–question sets, self-tests in chat, and individualized practice materials, and could personalize the content itself to learners’ prompts or preferences (e.g., Chun et al., 2025; Gao et al., 2025; Mzwri & Turcsányi-Szabo, 2025). In contrast, AI systems more often selected and delivered items from a question bank and could provide adaptive sequencing for subsequent questions based on observed performance and confidence ratings (e.g., Vertemati et al., 2024; Yan et al., 2024; Yin et al., 2025). Overall, GenAI’s adaptivity lies in the flexible generation of tailored test content, whereas AI affords data-driven selection and sequencing.

Learning management support occurred during goals and plans and enactment stages, and was evident in both technologies (AI: n = 4; GenAI: n = 2). In general, both AI and GenAI supported task announcements and reminders. However, AI further supported test and session scheduling, registration, and assignment submission and downloading (e.g., Al-Abdullatif et al., 2023; Hew et al., 2023), and GenAI enabled messages to be delivered in a more human-like and learner-friendly way (H. Li, 2023; Tsai et al., 2025).

Automated metacognitive report occurred during adaptation for reflection and subsequent adjustment, but remained limited in both technology groups (AI: n = 3; GenAI: n = 3). In this category, GenAI- and AI-based interventions were similar in that they translated activity and performance traces (e.g., grades, learning objectives, participation analyses, teachers’ feedback, and writing performance) into identified strengths, weaknesses, and specific barriers, and self-regulation reports (e.g., GenAI: Jin et al., 2025; Z. M. Liu et al., 2024; AI: Liao et al., 2024; Su, 2020). However, GenAI interventions could further extend reporting to the communicative and social layers, such as publishing participation summaries in a student group chat (H. Li, 2023).

In addition to the supports described above, AI and GenAI offered other distinct supports. AI provided negotiated help-seeking regulation (n = 1), which proposed or rejected help requests, and dynamically adjusted assistance levels (Chou et al., 2018). Additionally, one AI system permitted students to guide teachable agents (n = 1; Matsuda et al., 2020). Figure 5 presents the frequencies of AI- and GenAI-based interventions coded under the Operations facet of the COPES model.

Figure 5.

Matrix dot plot: AI- and GenAI-based SRL scaffolding in operations

AI- and GenAI-Based SRL Scaffolding in Products

Most existing AI- and GenAI-based systems support learners in producing outputs independently, rather than generating products in place of students. Within the Products facet, both technologies mainly focused on learning path products (AI: n = 2; GenAI: n = 3), which generate and dynamically update personalized learning paths over time. AI systems used learners’ goals, confidence, and performance self-assessments to arrange activity sequencing and recommendations and to update learner profiles (Afzaal et al., 2024; Carlon & Cross, 2022). Meanwhile, GenAI, in addition to using sub-goals and plans to generate and update learning paths (M. Zhu, 2025), could highlight key content from uploaded materials and mark levels of importance (i.e., high or low; Gao et al., 2025). Notably, assignment product generation was only afforded by GenAI (n = 2), which could generate draft tasks (e.g., essays) and revise them based on learners’ monitoring (e.g., Huang & Lin, 2024). However, it is important to acknowledge concerns regarding academic integrity and overreliance, and thus, careful scaffolding that promotes learning-before-automation is necessary.

AI- and GenAI-Based SRL Scaffolding in Evaluations

Grading students’ work was more frequently supported by AI than by GenAI in the included studies (AI: n = 6; GenAI: n = 1). AI systems scored language writing based on rubrics or predefined thresholds, or performance in SRL components (e.g., C. Liu et al., 2023; McCarthy et al., 2018; Weber et al., 2025). Meanwhile, only one GenAI system was used to grade students’ answers to English reading comprehension questions (Mzwri & Turcsányi-Szabo, 2025), as most GenAI-based designs concentrated on delivering qualitative feedback.

Cognitive feedback was utilized with comparable frequency between the reviewed AI and GenAI systems (AI: n = 14; GenAI: n = 13). In both groups, interventions commonly detected errors or misconceptions in learners’ work (e.g., drafts, answers, self-test performance, and goals) and supported improvement through hints and actionable next steps. However, they differed in interaction styles. GenAI feedback was more conversational and human-like, especially in language learning and writing, where it supported repeated drafting and revision through rubric-based or open-ended guidance (e.g., Campos, 2025; Mills et al., 2025; Mzwri & Turcsányi-Szabo, 2025). By contrast, AI feedback, grounded in predefined rules, was usually structured, with multi-level corrective feedback, color-coded annotations, or benchmark-based feedback (Perikos et al., 2017; Shafiee Rad, 2025; Weber et al., 2025). In short, GenAI tended to provide dialogue-based and generative feedback, whereas AI feedback was more structured and system-guided.

Metacognitive feedback appeared more frequently in reviewed AI-based systems than in GenAI (AI: n = 15; GenAI: n = 2). In both groups, the feedback focused on learners’ strategies and next-step guidance during task performance, but the way this support was generated differed. AI systems usually took in information on learners’ progress, goal-progress gaps, learning behaviors, or survey results, and used predefined rules to provide structured metacognitive suggestions (e.g., Chu et al., 2023; Serrano et al., 2018; Yan et al., 2024). By contrast, GenAI systems relied more on their generative ability, guided by system prompts, to interpret learners’ responses and offer contextual hints, reflective questions, and strategy support through dialogue (M.-R. Chen, 2024; Sun et al., 2025). Overall, metacognitive feedback was more common in AI systems and less common than cognitive feedback in both groups, which suggests that current systems have focused more on supporting task completion than on fostering metacognitive regulation.

Affective feedback in the reviewed studies for both AI and GenAI was limited (AI: n = 2; GenAI: n = 2). In the AI studies, systems often used emotion detection or learner-state monitoring to provide encouragement, supportive messages, or alternative tasks when learners seemed anxious or at risk of disengaging (H.-L. Liu et al., 2022; Mejeh & Rehm, 2024). Meanwhile, in the GenAI studies, the support was usually simpler, such as offering friendly encouragement during progress reflection or softening critical feedback with emotional support (e.g., K. Li et al., 2025). Overall, affective feedback in existing AI systems was more adaptive to learners’ states, while that of GenAI was more conversational and generally supportive. Figure 6 presents the frequencies of AI- and GenAI-based interventions coded under the Evaluations facet of the COPES model.

Figure 6.

Matrix dot plot: AI- and GenAI-based SRL scaffolding in evaluations

AI- and GenAI-Based SRL Scaffolding in Standards

In the Standards facet, only AI-based systems proactively addressed standards, whereas in GenAI systems, standards were either human-defined or randomly generated. Specifically, benchmark setting (n = 2) involved identifying essential tasks and required performance thresholds for each activity (e.g., McCarthy et al., 2018). Task prioritization (n = 1) arranged tasks in a sequence that best supports goal attainment (Afzaal et al., 2024).

Summary

In sum, the study revealed 19 distinct intervention types across included articles on AI- and GenAI-based SRL support systems. Collectively, both technology groups offered cognitive, metacognitive, and affective scaffolding; Q&A and learning content generation; test and exercise generation; learning management support; automated metacognitive reports; and learning path generation and updating, alongside evaluation functions, including grading and cognitive, metacognitive, and affective feedback. However, the distribution differed across groups. GenAI was primarily seen in Q&A and learning content generation, and in cognitive scaffolding and feedback, whereas the reviewed AI-based systems more strongly emphasized metacognitive scaffolding and feedback. Affective scaffolding and feedback, however, remained scarce, indicating an existing gap in supporting learners’ affective regulation in both AI- and GenAI-based SRL systems.

Beyond shared supports, each technology also afforded distinct functions. Particularly, AI evidenced negotiated help-seeking regulation and teachable-agent designs, and it proactively operationalized Standards during learning (e.g., setting benchmarks or prioritizing tasks based on learners’ goals). GenAI in the reviewed literature, by contrast, more clearly extended into the Products facet, particularly through assignment product generation and revision in collaboration with learners. However, most systems in both groups primarily aimed to scaffold learners to produce outputs independently, highlighting the importance of “self-doing” safeguards in current designs adopting generative tools.

From a COPES perspective, both AI- and GenAI-based SRL supports were concentrated in the Operations and Evaluations facets. The interventions were also predominantly situated in the enactment stage, with comparatively limited attention to task definition, goals and plans, and adaptation. This pattern indicated that current systems in both groups are still partially implementing the SRL process with an enactment bias that privileges in-task operational assistance and performance appraisal over preparatory and end-of-cycle reflective supports. The findings were aligned with prior review evidence. In particular, the forethought and self-reflection phases received relatively little attention in computer-assisted SRL supports (Prasse et al., 2024). In AI-based interventions, most studies focused on only one SRL phase (Chang & Sun, 2024), and in GenAI studies, performance and self-reflection were the most commonly targeted phases (Xia et al., 2026). Taken together, these findings suggest that partial SRL implementation is common in existing SRL scaffolding designs across different technologies. Therefore, the importance of designing for the full SRL cycle was further highlighted.

The two technology groups showed divergences in intervention approaches and formats. GenAI, grounded in generative capacities, tended to deliver operational supports and feedback through dialogic, human-like, and context-sensitive interaction to prompt actions and provide knowledge elaboration, iterative clarification, and revision guidance. Accordingly, GenAI appeared more capable of providing cognitive supports. Meanwhile, AI relied on predefined logic, structured interfaces, or learner-state modeling to provide analytics-triggered prompts, structured progress monitoring, and performance updates and reports, and was therefore more prevalently leveraged for metacognitive support. In the context where AI has continued to blossom alongside the growing applications of GenAI rather than being replaced, these differences allow the two technologies to functionally complement each other across different components, domains, and phases of the SRL architecture.

Taken together, these findings indicate convergent coverage of COPES facets, but divergent mechanisms (i.e., dialogic scaffolding vs. analytic scaffolding) between AI- and GenAI-based SRL systems. The study also reveals underdeveloped opportunities in affective supports, Standards settings, and SRL phases beyond enactment. Table 5 summarizes the coded interventions and their mapping to COPES facets and SRL stages. Full literature citations for the Research Question 2 findings are provided in Appendix 3.

Table 5.

Coded AI- and GenAI-Based SRL Interventions by COPES Facets and SRL Stages

Components/Stages	Stage 1: Task definition	Stage 2: Goals & plans	Stage 3: Enactment	Stage 4: Adaptation
Conditions	Source of Knowledge (A = 1 \| G = 4) ^(S)	Source of Knowledge (A = 2 \| G = 3) ^(S)	Source of Knowledge (A = 23 \| G = 33) ^(S)	Source of Knowledge (A = 3 \| G = 4) ^(S)
Conditions	Source of Instructions (A = 2 \| G = 1) ^(S)	Source of Instructions (A = 6 \| G = 6) ^(S)	Source of Instructions (A = 34 \| G = 22) ^(S)	Source of Instructions (A = 5 \| G = 4) ^(S)
Operations	Cognitive Scaffolding (A = 3 \| G = 4) ^(S) Q&A and Learning Content Generation (G = 1) ^(G)	Cognitive Scaffolding (A = 1) ^(A) Learning Management (A = 2 \| G = 2) ^(S) Metacognitive Scaffolding (A = 5 \| G = 4) ^(S)	Affective Scaffolding (A = 1 \| G = 1) ^(S) Cognitive Scaffolding (A = 9 \| G = 13) ^(S) Learning Management Support (A = 2) ^(A) Metacognitive Scaffolding (A = 15 \| G = 5) ^(S) Negotiated Help-seeking Regulation (A = 1) ^(A) Q&A and Learning Content Generation (A = 3 \| G = 18) ^(S) Teachable Agent (A = 1) ^(A) Test & Exercise Generation (A = 5 \| G = 6) ^(S)	Affective Scaffolding (A = 2) ^(A) Metacognitive Scaffolding (A = 4 \| G = 2) ^(S) Test & Exercise Generation (G = 1) ^(G) Automated Metacognitive Report (A = 3 \| G = 3) ^(S)
Products		Learning Path Products (A = 1 \| G = 2) ^(S)	Assignment Product Generation (G = 2) ^(G)	Learning Path Products (A = 1 \| G = 1) ^(S)
Evaluations			Affective feedback (A = 2 \| G = 2) ^(S)
			Cognitive Feedback (A = 14 \| G = 13) ^(S)
			Grading (A = 6 \| G = 1) ^(S)
			Metacognitive Feedback (A = 15 \| G = 2) ^(S)
Standards		Task Prioritization (A = 1) ^(A)
Standards		Benchmark Setting (A = 2) ^(A)

Note. A = AI; G = GenAI; numbers indicate frequencies. ^(A) = AI only; ^(G) = GenAI only; ^(S) = Shared. Counts for “source of knowledge” and “source of instructions” may differ from counts in subsequent COPES facets because a single intervention can serve both functions, and multiple interventions can simultaneously function as sources of instruction.

RQ3. How can GenAI and AI More Comprehensively Support the SRL Process?

Based on findings from both pedagogical and technological perspectives, this section proposes a framework for a comprehensive SRL design argument that includes potential interventions across the entire SRL cycle, with system adaptivity and learner adaptability, and the consideration of the complementary strengths of AI and GenAI (Figure 7).

Figure 7.

A framework for AI-, GenAI-, and hybrid SRL scaffolding

Key Design Elements for Full-Cycle SRL

First, we contend that the inclusion of affective supports is important, besides the dominant presence of cognitive and metacognitive supports. Panadero (2017) describes SRL as involving cognitive, behavioral, metacognitive, motivational, and emotional aspects. To date, AI- and GenAI-based supports have disproportionately targeted cognitive and metacognitive processes, while affective support has been underutilized. This concern is consistent with Banihashem et al. (2025). Winne (2018) emphasized motivation and emotion in the SRL process, and argued that they shape learners’ goals and influence how they monitor their learning. In this sense, affective support may help learners better manage their motivation and emotions and remain engaged in learning. Consistent with this view, some primary studies have reported that affective support increase engagement and positive emotions, including motivation and enjoyment, while reducing negative feelings, such as anxiety and isolation (Daradoumis & Arguedas, 2020; Hew et al., 2023). Based on the existing evidence, future designs can either use emotion and state sensing and monitoring to trigger adaptive support (with the integration of one or more technologies such as NLP, fuzzy logic, machine learning, and computer vision) or emotion-support messages (through prompt engineering with GenAI).

Second, support for environmental and resource management should be expanded. From a theoretical perspective, the Conditions in the COPES model refer to external resources and environmental features that shape what learners can do in subsequent processes. In this sense, relevant supports can assist students in locating and mobilizing time, resources, and social context. They also align closely with help-seeking and environmental structuring in Zimmerman’s (2000) SRL discussion, and resource management in Pintrich’s (2000) framework. Given that, we contend that SRL systems can incorporate domain-specific content generation and learning management support. In particular, GenAI-based systems, when integrated with RAG, provide timely, discipline-specific resources. Also, learning management systems may support scheduling and daily task reminders throughout the four SRL stages. However, these supports were more often embedded within broader frameworks that included cognitive scaffolds. In these studies, benefits were reported for students’ personalized learning, SRL practices, autonomy, engagement, and the use of behavioral strategies (Chun et al., 2025; Pan et al., 2025).

Third, the inclusion of Standards should be explicitly expanded. Standards in the COPES serve as criteria for evaluating performance across SRL stages, but are largely neglected in existing designs. We argue that students’ SRL can be strengthened through explicit, negotiable standards, that is, benchmarks, success criteria, and task priorities. The standards can enable students to proactively manage their performance and dynamically negotiate as learning needs evolve, for example, an existing AI system based on learners’ goals to identify the following tasks with required scores and task priority (Afzaal et al., 2024). However, our findings show that system-driven Standards were found only in AI-based systems, whereas standards in GenAI systems were often human-fed rubrics or loosely specified. We envision that future systems can further tailor standards to learners’ profiles, past performance, and prior knowledge.

Fourth, we propose incorporating visualization as a core design feature of SRL support systems. Importantly, visualization is not a standalone intervention; rather, it serves as a delivery mechanism that enhances interpretability and reduces unnecessary cognitive load. This is consistent with Mayer’s (2021) cognitive theory of multimedia learning, which emphasizes that well-designed visual representations can reduce extraneous cognitive load and facilitate active processing of selecting, organizing, and integrating information. In the included studies, visualization was predominantly learning-analytic-based and was frequently used to support learning path generation and revision, knowledge maps and cards, task progress updates, and metacognitive reports. These applications supported Operations (e.g., monitoring and translating), Products (e.g., visualizing learning paths), and Evaluations (e.g., reporting task progress), and extended across goals and plans, enactment, and adaptation. Future designs can expand visualization into task-definition processes, for example, by generating a map of key task concepts and explicitly contrasting them with learners’ prior knowledge indicators, thereby helping learners clarify task demands and plan more effectively before setting goals and engaging in task performance.

Finally, the collection of the above recommended activities can be designed within an overall full SRL cycle. Supporting the broader SRL cycle is not only theoretical but also important in practice. First, the cyclical nature of the SRL process has been well conceptualized in established SRL models (Pintrich, 2000; Winne & Hadwin, 1998; Zimmerman, 2000). However, the findings of the present study show that GenAI and AI supports were predominantly promoted during the enactment (performance) stage, with less focus on the other stages of task definition, goals and plans, and adaptation. As a result, such designs may strengthen immediate task performance and performance-related regulations, such as self-monitoring and help-seeking, but may limit learners’ long-term ability to independently and effectively regulate future learning across preparation, performance, and post-task adjustment. This concern was consistent with previous review evidence. In particular, only a minority of studies incorporated GenAI into the full SRL cycle, while the remaining studies either treated SRL as a single skill without phase division or addressed only one or two phases (Xia et al., 2026). In the broader context of computer-assisted SRL, Prasse et al. (2024) showed that full-SRL-cycle studies were limited, and recommended a comprehensive approach in future designs.

However, each intervention does not necessarily need to be distributed equally across all SRL phases. Rather, different functions may be more strongly positioned in particular phases. Our findings showed that cognitive, metacognitive, and affective supports spanned four SRL stages, whereas learning management supports were scattered across goals and plans, and enactment. Content generation is most concentrated in enactment, and standards setting is located in goals and plans. Taken together, this study recommends a design framework in which SRL-supporting technologies (GenAI, AI, and hybrid) collectively contribute to the full SRL cycle according to their pedagogical relevance.

Adaptivity and Adaptability

We propose two complementary approaches to AI- and GenAI-based SRL scaffolding. System adaptivity refers to a system’s ability to dynamically adjust to the current task and learner progress (Fischer, 2001). In contrast, learner adaptability denotes learners’ control over the instructional support and content of the learning environment (Fischer, 2001).

System Adaptivity

Adaptivity is important, as it enhances students’ performance in the system. This was evident in a prior review that showed that AI adaptivity imposes medium-to-large effects on learners’ cognitive learning outcomes (X. Wang et al., 2024). In the included studies, adaptivity was observed in both technologies across COPES facets. Particularly, within Operations, systems provided real-time next-step hints, adaptive tests and content generation, and dynamic knowledge maps. Those in other facets included dynamic, personalized learning paths within Products, real-time contextual feedback within Evaluations, and adaptive benchmarks and task prioritization within Standards. In future adaptivity, scholars recommend using prior-knowledge pretests or learner profiles as a basis to customize the learning process and chatbot interactions (Gao et al., 2025; M. Liu & Reinders, 2025).

Learner Adaptability

Learner adaptability positions learners as regulators who actively accept or resist external assistance and purposefully control outputs across the SRL process. We contend that this dimension is necessary given issues of overreliance on external support, for example, overreliance on GenAI-generated feedback (e.g., Campos, 2025), the tendency of lower achievers to make immediate requests for GenAI assistance (Xu et al., 2024), and feeling overwhelmed with AI metacognitive prompts (McCarthy et al., 2018). Consequently, such issues may foster immediate task completion, but reduce the meaning of systems’ support to learners’ needs and learners’ longer-term capacity to regulate their own learning. Several supports enhancing students’ adaptability were recorded in several systems, for example, offering regulatory choices over difficulty or learning levels, implementing reward-punishment systems to encourage or reduce certain learning behaviors, and allowing learners to filter with preferences (e.g., language proficiency, length, and complexity; Chun et al., 2025; Hwang et al., 2025; Pan et al., 2025; Xu et al., 2024; Yan et al., 2024).

In conclusion, our study underscores the importance of aligning AI- and GenAI-based SRL interventions with pedagogical goals. Figure 7 illustrates the proposed conceptual framework, in which AI-, GenAI-, and hybrid supports collectively contribute to the four SRL stages: task definition, goals and plans, enactment, and adaptation. The recommended functions included cognitive, metacognitive, and affective supports; learning management support; content generation; and learning standards. These supports are shown as being applicable across the SRL cycle. The framework recommends using visualizations as design features. The framework highlights adaptivity and adaptability to support personalized learning.

Conclusion, Implications, and Limitations

Conclusions

Continuing the work of previous studies (Banihashem et al., 2025; Chang et al., 2023; Chang & Sun, 2024; Sardi et al., 2025), this study offers an evidence-based synthesis of 70 empirical articles published between 2015 and 2025, which shows how GenAI mechanisms relate to or extend beyond those of earlier AI approaches in supporting SRL. This systematic review focuses on design taxonomy with explicit identification of technological foundations and pedagogical applications. We mapped our review onto Winne and Hadwin’s (1998) COPES model to emphasize purposeful design. Ultimately, we introduce an actionable framework for AI- and GenAI-based SRL scaffolding.

Regarding technological employment, we found that existing AI-based SRL interventions primarily employed knowledge-based systems, machine learning (ML), and natural language processing (NLP), supplemented by uncertainty and probabilistic reasoning, search and optimization, multi-agent systems, and computer vision. GenAI predominantly utilized text-based LLMs, accessed through APIs, with prompt engineering, typically without knowledge bases. Although their mechanisms differ, both GenAI and AI shared the goal of providing personalized and adaptive learning processes.

From a pedagogical perspective, both AI and GenAI provided cognitive, metacognitive, and affective supports; learning content and assessments; task progress; learning administrative support; prompts for self-reflection; and automated metacognitive reports. Among those interventions, affective scaffolding and feedback remained scarce in both technologies, indicating a persistent gap in supporting learners’ affective regulation. Across the COPES facets and SRL stages, interventions primarily targeted Operations and Evaluations in the enactment phase and were less focused on other phases of task definition, goal setting, or adaptation. This reflects an enactment bias in current SRL scaffolding designs that especially promotes in-task operational assistance and evaluation. Regarding differences between the two technology groups, AI in the reviewed literature was utilized more for structured metacognitive supports and system-driven setting of Standards (e.g., benchmark setting and task prioritization). Included GenAI-based systems, however, suggested distinct effectiveness in cognitive scaffolding with context-aware dialogue and domain-adaptive content generation. Taken together, these findings suggest convergence in pedagogical coverage but divergence in scaffolding mechanisms between AI and GenAI, and also reveal underdeveloped opportunities in affective supports, Standards setting, and SRL phases beyond enactment.

Finally, this study posits that a comprehensive SRL support system should include key components of cognitive, metacognitive, and affective support; domain-specific content generation; learning administrative support; standards setting; and visualizations as an additional delivery mechanism. These supports are shown as applicable across the SRL cycle. Importantly, we underscore adaptivity and adaptability as essential components of personalization: adaptivity affords effective and timely SRL scaffolding, while adaptability enables students to control human-system interactions to meet their ongoing needs.

Implications for Research and Practice

This study aimed to deepen both theoretical and practical understandings of AI and GenAI in supporting SRL. Theoretically, this review contributes to the literature by providing deeper insights into specific AI and GenAI technologies deployed for SRL purposes, their pedagogical applications, and, ultimately, a comprehensive SRL scaffolding framework.

Practically, the proposed framework provides a comprehensive SRL design for instructional designers and educators to develop their own SRL systems using GenAI, AI, or hybrid approaches. In general, aligned with previous studies, we contend that the scaffolding design should address the full SRL cycle, ranging from supports on task analysis, goal setting, task implementation, and post-task reflection and adaptation. The following part presents several further implications that emerged from the present review.

First, we found that current systems of both technologies prefer cognitive and metacognitive over affective support, consistent with previous SRL reviews. Future studies should treat affective support as a core design element to help learners regulate their frustration, confidence, and persistence through adaptive prompts, reflective check-ins, or emotionally responsive conversational scaffolds.

Second, standards are part of COPES but are largely neglected in the current SRL system design. Future systems should make standards (e.g., benchmarks, criteria, and task priority) explicit, visible, and negotiable for students.

Third, visualization (e.g., progress, gaps, learning paths, and task structures) can serve as a delivery mechanism, improving self-monitoring and knowledge organization, reducing extraneous cognitive load, and supporting adaptation.

Fourth, GenAI has served as a highly flexible and responsive resource for learning content and help-seeking. However, learning designers should set “self-doing before auto-generation” rules and acknowledge the ethical concerns and risks of overreliance when prompting systems to generate students’ work.

Fifth, neither AI nor GenAI is better overall; rather, it is a matter of which technology is better suited to the intended SRL support. For example, if we aim to track progress or provide analytics-triggered metacognitive feedback, earlier AI may be especially suitable. When it comes to dialogic explanation and elaboration, GenAI can be an effective option. Several studies have piloted the use of GenAI in the communicative and social layers for analytics-driven “back-end” processes (e.g., H. Li, 2023), suggesting a promising direction for future hybrid design.

Lastly, personalization in GenAI-, AI-, and hybrid systems should involve both system adaptivity and learner adaptability. Future empirical studies may usefully compare the effects of system-adaptive-only designs with system-adaptive-plus-learner-controllable designs.

Limitations

This systematic review has several limitations. First, regarding the technological reports, five of the 70 articles did not specify the AI and GenAI technologies used, which limited our analysis. Second, the counts presented in the results and discussion sections of RQ1 and RQ2 were intended to be exploratory and descriptive and should not be interpreted as effect-size comparisons. Moreover, regarding technical coding, there may occasionally be minor technical specifications of the systems that were not explicitly reported, so we acknowledge that the study’s frequency analysis may not reflect the full technological landscape. Furthermore, studies that include multiple interventions or use the same intervention across several SRL phases may carry greater weight in the frequency analysis. In this sense, the frequency analysis should be interpreted as an indicator of patterns and trends in intervention types within existing SRL systems, rather than as a direct measurement of impact. Third, although the comparison of the AI and GenAI groups was feasible, broader evidence from GenAI studies over time would strengthen the observed patterns. Fourth, we encountered challenges in stage 4 (adaptation) of Winne and Hadwin’s (1998) SRL model. As Greene and Azevedo (2007) noted, it is unclear how to distinguish between long-term changes and updates in earlier phases, particularly in stage 3. Multiple authors have sought to reach consensus on the subcodes for stages 3 (evaluations) and 4. Nevertheless, Winne and Hadwin’s model remains distinct in examining how external supports (with AI and GenAI) can support the dynamic and iterative SRL process. Lastly, this review focused on the design taxonomy of AI- and GenAI-based SRL interventions (technological foundations and pedagogical applications). Future research could employ meta-analysis methods to investigate the effects of the interventions on learning outcomes.

Footnotes

Acknowledgements

We sincerely thank Dr. Abebayehu Yohannes for his valuable support and constructive feedback throughout the project development.

ORCID iDs

Hue Nguyen

Hsiu-Ling Chen

Khoirudin Asfani

Ethical Considerations

Secondary analysis of published articles; no human participants or identifiable data. Ethics approval and consent not required.

Authors’ Contributions

Hue Nguyen led framework development, data collection and analysis, visualization, and preparation of the initial manuscript draft. Hsiu-Ling Chen contributed to framework development and revised subsequent drafts. Khoirudin Asfani participated in data collection and analysis and revised subsequent drafts.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by National Science and Technology Council [Grant number: NSTC 113-2410-H-011 -003 -MY3], the “Empower Vocational Education Research Center” of National Taiwan University of Science and Technology (NTUST) from the Featured Areas Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education (MOE) in Taiwan.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

All data analyzed during this study are included in this published article and its supplementary information files.*

Author Biographies

Hue Nguyen is a PhD candidate at the Graduate Institute of Digital Learning and Education, National Taiwan University of Science and Technology, Taipei, Taiwan. Her academic interest areas are the instructional design of online learning, self-regulated learning (SRL), and AI-based chatbots. Over the past four years, she has researched with a focus on SRL, Discussion Forums, and the physical learning environment in online learning. Email: D11211817@mail.ntust.edu.tw.

Hsiu-Ling Chen is a Professor of Digital Learning and Education at the National Taiwan University of Science and Technology. Her main research interest is ICT integration in education, including flipped classrooms, CSCL (computer-supported collaborative learning), digital storytelling, and high-order cognitive thinking (critical thinking, creative thinking, problem-solving, metacognition).

Khoirudin Asfani is a doctoral student at the Graduate Institute of Digital Learning and Education, National Taiwan University of Science and Technology, Taiwan. His research interests are ICT integration in education, including computer-supported collaborative learning in computer education, higher-order thinking skills (computational thinking, problem solving, creative thinking, critical thinking), and informatics engineering education. Email: khoirudin.asfani.ft@um.ac.id.

Appendix

References

*Asterisk: included studies for the current review.

* Afzaal

Zia

Nouri

Fors

(2024). Informative feedback and explainable AI-Based recommendations to support students’ self-regulation. Technology, Knowledge and Learning, 29(1), 331–354. https://doi.org/10.1007/s10758-023-09650-0

* Al-Abdullatif

A. M.

Al-Dokhny

A. A.

Drwish

A. M.

(2023). Implementing the Bashayer chatbot in Saudi higher education: Measuring the influence on students' motivation and learning strategies. Frontiers in Psychology, 14(1), Article 1129070. https://doi.org/10.3389/fpsyg.2023.1129070

* Al-Shanfari

Demmans Epp

Baber

Nazir

(2020). Visualising alignment to support students’ judgment of confidence in open learner models. User Modeling and User-Adapted Interaction, 30(1), 159–194. https://doi.org/10.1007/s11257-019-09253-4

* Arvidsson

T. S.

Kuhn

(2021). Realizing the full potential of individualizing learning. Contemporary Educational Psychology, 65(1), Article 101960. https://doi.org/10.1016/j.cedpsych.2021.101960

Asfani

Chen

H.-L.

(2025). Problem or project-based computer-supported collaborative learning practices in computer education: A systematic review of SSCI articles published from 2014 to 2023. Education and Information Technologies, 30(6), 7627–7658. https://doi.org/10.1007/s10639-024-13125-9

Banh

Strobel

(2023). Generative artificial intelligence. Electronic Markets, 33(1), 63. https://doi.org/10.1007/s12525-023-00680-1

Banihashem

S. K.

Bond

Bergdahl

Khosravi

Noroozi

(2025). A systematic mapping review at the intersection of artificial intelligence and self-regulated learning. International Journal of Educational Technology in Higher Education, 22(1), 50. https://doi.org/10.1186/s41239-025-00548-8

* Campos

(2025). AI-assisted feedback in CLIL courses as a self-regulated language learning mechanism: Students’ perceptions and experiences. European Public & Social Innovation Review, 10(1), 1–14. https://doi.org/10.31637/epsir-2025-1568

10.

* Carlon

M. K. J.

Cross

J. S.

(2022). Knowledge tracing for adaptive learning in a metacognitive tutor. Open Education Studies, 4(1), 206–224. https://doi.org/10.1515/edu-2022-0013

11.

Chang

D. H.

Lin

M. P.-C.

Hajian

Wang

Q. Q.

(2023). Educational design principles of using AI chatbot that supports self-regulated learning in education: Goal setting, feedback, and personalization. Sustainability, 15(17), Article 12921. https://doi.org/10.3390/su151712921

12.

Chang

W.-L.

Sun

J. C.-Y.

(2024). Evaluating AI's impact on self-regulated language learning: A systematic review. System, 126(1), Article 103484. https://doi.org/10.1016/j.system.2024.103484

13.

* Chen

M.-R.

(2024). Metacognitive mastery: Transformative learning in EFL through a generative AI chatbot fueled by metalinguistic guidance. Educational Technology & Society, 27(3), 407–427. https://doi.org/10.30191/ETS.202407_27(3).TP05

14.

* Chen

Xiang

Zhou

Jia

Shang

Gasevic

Fan

(2025). Unpacking help-seeking process through multimodal learning analytics: A comparative study of ChatGPT vs human expert. Computers & Education, 226(1), Article 105198. https://doi.org/10.1016/j.compedu.2024.105198

15.

* Chou

C.-Y.

Lai

K. R.

Chao

P.-Y.

Tseng

S.-F.

Liao

T.-Y.

(2018). A negotiation-based adaptive learning system for regulating help-seeking behaviors. Computers & Education, 126(1), 115–128. https://doi.org/10.1016/j.compedu.2018.07.010

16.

* Chu

S.-T.

Hwang

G.-J.

Chien

S.-Y.

Chang

S.-C.

(2023). Incorporating teacher intelligence into digital games: An expert system‐guided self‐regulated learning approach to promoting EFL students' performance in digital gaming contexts. British Journal of Educational Technology, 54(2), 534–553. https://doi.org/10.1111/bjet.13260

17.

* Chun

Kim

Lee

Cho

Kim

Chung

Heo

(2025). A comparative analysis of On-Device AI-Driven, self-regulated learning and traditional pedagogy in University Health Sciences education. Applied Sciences, 15(4), 1815. https://doi.org/10.3390/app15041815

18.

* Daradoumis

Arguedas

(2020). Cultivating students’ reflective learning in metacognitive activities through an affective pedagogical agent. Educational Technology & Society, 23(2), 19–31. https://doi.org/10.30191/ETS.202004_23(2).0002

19.

* Duffy

M. C.

Azevedo

(2015). Motivation matters: Interactions between achievement goals and agent scaffolding for self-regulated learning within an intelligent tutoring system. Computers in Human Behavior, 52(1), 338–348. https://doi.org/10.1016/j.chb.2015.05.041

20.

* Du

Hew

K. F.

Zhang

(2025). Designing a recommender system to promote self-regulated learning in online contexts: A design-based study. Education and Information Technologies, 30(2), 2389–2428. https://doi.org/10.1007/s10639-024-12867-w

21.

Durak

Çankaya

Özdemir

Can

(2024). Artificial intelligence in education: A bibliometric study on its role in transforming teaching and learning. International Review of Research in Open and Distance Learning, 25(3), 219–244. https://doi.org/10.19173/irrodl.v25i3.7757

22.

Elo

Kyngäs

(2008). The qualitative content analysis process. Journal of Advanced Nursing, 62(1), 107–115. https://doi.org/10.1111/j.1365-2648.2007.04569.x

23.

* Ferguson

van den Broek

E. L.

van Oostendorp

(2022). AI-Induced guidance: Preserving the optimal Zone of Proximal Development. Computers and Education: Artificial Intelligence, 3(1), Article 100089. https://doi.org/10.1016/j.caeai.2022.100089

24.

Fischer

(2001). User modeling in human–computer interaction. User Modeling and User-Adapted Interaction, 11(1), 65–86. https://doi.org/10.1023/A:1011145532042

25.

* Gao

Shao

Lin

Yue

Ieong

Sun

Zauner

Wei

Chen

(2025). Fine-tuned large language model for visualization system: A study on self-regulated learning in education. IEEE Transactions on Visualization and Computer Graphics, 31(1), 514–524. https://doi.org/10.1109/TVCG.2024.3456145

26.

* Glick

Miedijensky

Zhang

(2024). Examining the effect of AI-powered virtual-human training on STEM majors’ self-regulated learning behavior. Frontiers in Education, 9(1), Article 1465207. https://doi.org/10.3389/feduc.2024.1465207

27.

Greene

J. A.

Azevedo

(2007). A theoretical review of Winne and Hadwin’s model of self-regulated learning: New perspectives and directions. Review of Educational Research, 77(3), 334–372. https://doi.org/10.3102/003465430303953

28.

* Guerrero-Roldán

A. E.

Rodríguez-González

M. E.

Bañeres

Elasri-Ejjaberi

Cortadas

(2021). Experiences in the use of an adaptive intelligent system to enhance online learners’ performance: A case study in Economics and Business courses. International Journal of Educational Technology in Higher Education, 18(1), 36. https://doi.org/10.1186/s41239-021-00271-0

29.

* Han

Jin

Choi

(2026). Mobile-based artificial intelligence chatbot for self-regulated learning in a hybrid flipped classroom. Journal of Computing in Higher Education, 38(1), 285–309. https://doi.org/10.1007/s12528-025-09434-8

30.

* Hew

Huang

Jia

(2023). Using chatbots to support student goal setting and social presence in fully online activities: Learner engagement and perceptions. Journal of Computing in Higher Education, 35(1), 40–68. https://doi.org/10.1007/s12528-022-09338-x

31.

Hong

Q. N.

Fàbregues

Bartlett

Boardman

Cargo

Dagenais

Gagnon

M.-P.

Griffiths

Nicolau

O’Cathain

Rousseau

M.-C.

Vedel

Pluye

(2018). The Mixed Methods Appraisal Tool (MMAT) version 2018 for information professionals and researchers. Education for Information, 34(4), 285–291. https://doi.org/10.3233/EFI-180221

32.

* Huang

Lin

H. C.

(2024). ChatGPT as a life coach for professional identity formation in medical education: A self-regulated learning perspective. Educational Technology & Society, 27(3), 374–389. https://doi.org/10.30191/ETS.202407_27(3).TP03.

33.

* Huang

A. Y. Q.

Lin

C.-Y.

S.-Y.

Yang

S. J. H.

(2025). The impact of GenAI-enabled coding hints on students’ programming performance and cognitive load in an SRL-based Python course. British Journal of Educational Technology, 56(5), 1942–1972. https://doi.org/10.1111/bjet.13589

34.

* Hu

Y.-H.

Hsieh

C.-L.

Salac

E. S. N.

(2024). Advancing freshman skills in information literacy and self-regulation: The role of AI learning companions and Mandala Chart in academic libraries. The Journal of Academic Librarianship, 50(3), Article 102885. https://doi.org/10.1016/j.acalib.2024.102885

35.

* Hwang

G.-J.

Cheng

P.-Y.

Chang

C.-Y.

(2025). Facilitating students’ critical thinking, metacognition, and problem-solving tendencies in geriatric nursing class: A mixed-method study. Nurse Education in Practice, 83(1), Article 104266. https://doi.org/10.1016/j.nepr.2025.104266

36.

* Jin

F. J. Y.

Maheshi

Lai

Gasevic

Chen

Charwat

Chan

P. W. K.

Martinez-Maldonado

Gašević

Tsai

Y.-S.

(2025). Students’ perceptions of generative AI-Powered learning analytics in the feedback process: A feedback literacy perspective. Journal of Learning Analytics, 12(1), 152–168. https://doi.org/10.18608/jla.2025.8609

37.

* Klar

(2025). Using ChatGPT is easy, using it effectively is tough? A mixed methods study on K-12 students’ perceptions, interaction patterns, and support for learning with generative AI chatbots. Smart Learning Environments, 12(1), 32. https://doi.org/10.1186/s40561-025-00385-2

38.

* Lai

J. W.

Qiu

Thway

Zhang

Jamil

N. B.

C. L.

S. S. H.

Lim

F. S.

(2025). Leveraging process-action epistemic network analysis to illuminate student self-regulated learning with a socratic chatbot. Journal of Learning Analytics, 12(1), 32–49. https://doi.org/10.18608/jla.2025.8549

39.

* Lee

H.-Y.

Chen

P.-H.

Wang

W.-S.

Huang

Y.-M.

T.-T.

(2024). Empowering ChatGPT with guidance mechanism in blended learning: Effect of self-regulated learning, higher-order thinking skills, and knowledge construction. International Journal of Educational Technology in Higher Education, 21(1), 16. https://doi.org/10.1186/s41239-024-00447-4

40.

* Li

(2023). Effects of a ChatGPT-based flipped learning guiding approach on learners’ courseware project performances and perceptions. Australasian Journal of Educational Technology, 39(5), 40–58. https://doi.org/10.14742/ajet.8923

41.

* Liao

Zhang

Wang

Luo

(2024). Design and implementation of an AI‐enabled visual report tool as formative assessment to promote learning achievement and self‐regulated learning: An experimental study. British Journal of Educational Technology, 55(3), 1253–1276. https://doi.org/10.1111/bjet.13424

42.

* Li

Lan

(2025). Comparative analysis of GPT-4.0 and teacher feedback on student-generated questions in the flipped classroom. Educational Technology Research & Development, 73(3), 1347–1369. https://doi.org/10.1007/s11423-025-10467-0

43.

* Lim

Bannert

van der Graaf

Fan

Rakovic

Singh

Molenaar

Gašević

(2024). How do students learn with real-time personalized scaffolds? British Journal of Educational Technology, 55(4), 1309–1327. https://doi.org/10.1111/bjet.13414

44.

* Lin

C.-J.

Wang

W.-S.

Lee

H.-Y.

P.-H.

Huang

T.-T.

(2025). Advancing self-directed learning in STEM education: Integrating GPT-based learning aid with multimodal learning analytics. Journal of Research on Technology in Education. Advance online publication. https://doi.org/10.1080/15391523.2025.2504358

45.

* Li

Sadiq

Qambar

Cheng

(2025). The impact of ChatGPT integrated instruction on students’ research skills: The mediating effects of autonomous motivation, engagement, and self-directed learning. Education and Information Technologies, 30(1), 13073. https://doi.org/10.1007/s10639-024-13303-9

46.

* Liu

Hou

Y. F.

Wang

Hwang

G. J.

(2023). Incorporating a reflective thinking promoting mechanism into artificial intelligence-supported English writing environments. Interactive Learning Environments, 31(9), 5614–5632. https://doi.org/10.1080/10494820.2021.2012812

47.

* Liu

Z. M.

Hwang

G. J.

Chen

C. Q.

Chen

X. D.

(2024b). Integrating large language models into EFL writing instruction: Effects on performance, self-regulated learning strategies, and motivation. Computer Assisted Language Learning. Advance online publication. https://doi.org/10.1080/09588221.2024.2389923

48.

* Liu

C.-C.

Hwang

G.-J.

Y.-F.

Wang

(2025). Effects of an automated corrective feedback-based peer assessment approach on students’ learning achievement, motivation, and self-regulated learning conceptions in foreign language pronunciation. Educational Technology Research and Development, 73(4), 2403–2424. https://doi.org/10.1007/s11423-025-10484-z

49.

* Liu

Reinders

(2025). Do AI chatbots impact motivation? Insights from a preliminary longitudinal study. System, 128(1), Article 103544. https://doi.org/10.1016/j.system.2024.103544

50.

* Liu

H.-L.

Wang

T.-H.

Lin

H.-CK.

Lai

C.-F.

Huang

Y.-M.

(2022). The influence of affective feedback adaptive learning system on learning engagement and self-directed learning. Frontiers in Psychology, 13(1), Article 858411. https://doi.org/10.3389/fpsyg.2022.858411

51.

* Liu

Zhang

Jia

(2024a). The design of guiding and adaptive prompts for intelligent tutoring systems and its effect on students’ mathematics learning. IEEE Transactions on Learning Technologies, 17(1), 1379–1389. https://doi.org/10.1109/TLT.2024.3382000

52.

* Long

Aleven

(2017). Enhancing learning outcomes through self-regulated learning support with an Open Learner Model. User Modeling and User-Adapted Interaction, 27(1), 55–88. https://doi.org/10.1007/s11257-016-9186-6

53.

* Matsuda

Weng

Wall

(2020). The effect of metacognitive scaffolding for learning by teaching a teachable agent. International Journal of Artificial Intelligence in Education, 30(1), 1–37. https://doi.org/10.1007/s40593-019-00190-2

54.

Mayer

R. E.

(2021). Cognitive theory of multimedia learning. In Mayer

R. E.

Fiorella

(Eds.), The Cambridge handbook of multimedia learning (3rd ed., pp. 57–72). Cambridge University Press. https://doi.org/10.1017/9781108894333.008

55.

* McCarthy

K. S.

Likens

A. D.

Johnson

A. M.

Guerrero

T. A.

McNamara

D. S.

(2018). Metacognitive overload!: Positive and negative effects of metacognitive prompts in an intelligent tutoring system. International Journal of Artificial Intelligence in Education, 28(3), 420–438. https://doi.org/10.1007/s40593-018-0164-5

56.

* Mejeh

Rehm

(2024). Taking adaptive learning in educational settings to the next level: Leveraging natural language processing for improved personalization. Educational Technology Research and Development, 72(3), 1597–1621. https://doi.org/10.1007/s11423-024-10345-1

57.

* Mills

Hok

Dressen

Veillas

(2025). The design and evaluation of an interactive AI companion for foreign language writing. Foreign Language Annals, 58(1), 40–69. https://doi.org/10.1111/flan.12790

58.

* Mzwri

Turcsányi-Szabo

(2025). The impact of prompt engineering and a generative AI-Driven tool on autonomous learning: A case Study. Education Sciences, 15(2), 199. https://doi.org/10.3390/educsci15020199

59.

* Neshaei

S. P.

Mejia‐Domenzain

Davis

Käser

(2025). Metacognition meets AI: Empowering reflective writing with large language models. British Journal of Educational Technology, 56(5), 1864–1896. https://doi.org/10.1111/bjet.13601

60.

* Neumann

Yin

Sowe

S. K.

Decker

Jarke

(2025). An LLM-driven chatbot in higher education for databases and information systems. IEEE Transactions on Education, 68(1), 103–116. https://doi.org/10.1109/TE.2024.3467912

61.

* Ng

D. T. K.

Tan

Leung

(2024). Empowering student self‐regulated learning and science education through ChatGPT: A pioneering pilot study. British Journal of Educational Technology, 55(4), 1328–1353. https://doi.org/10.1111/bjet.13454

62.

* Nguyen

Gul

Dang

Huynh

Tuunanen

(2025). Designing embodied generative artificial intelligence in mixed reality for active learning in higher education. Innovations in Education & Teaching International, 62(5), 1632–1647. https://doi.org/10.1080/14703297.2025.2499177

63.

Ohm

(2024). Focusing on Fine-Tuning: Understanding the four pathways for shaping generative AI. The Columbia Science & Technology Law Review, 25(2), 214. https://doi.org/10.52214/stlr.v25i2.12762

64.

* Ortega-Ochoa

Quiroga Pérez

Arguedas

Daradoumis

Marquès Puig

J. M.

(2024). The effectiveness of empathic chatbot feedback for developing computer competencies, motivation, self-regulation, and metacognitive reasoning in online higher education. Internet of Things, 25(1), Article 101101. https://doi.org/10.1016/j.iot.2024.101101

65.

Page

M. J.

McKenzie

J. E.

Bossuyt

P. M.

Boutron

Hoffmann

T. C.

Mulrow

C. D.

Shamseer

Tetzlaff

J. M.

Akl

E. A.

Brennan

S. E.

Chou

Glanville

Grimshaw

J. M.

Hróbjartsson

Lalu

M. M.

Loder

E. W.

Mayo-Wilson

McDonald

Moher

(2021). The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. PLoS Medicine, 18(3), Article e1003583. https://doi.org/10.1371/journal.pmed.1003583

66.

Panadero

(2017). A review of self-regulated learning: Six models and four directions for research. Frontiers in Psychology, 8(1), 422. https://doi.org/10.3389/fpsyg.2017.00422

67.

Panadero

Klug

Järvelä

(2016). Third wave of measurement in the self-regulated learning field: When measurement and intervention come hand in hand. Scandinavian Journal of Educational Research, 60(6), 723–735. https://doi.org/10.1080/00313831.2015.1066436

68.

* Pan

Lai

Guo

(2025). Effects of GenAI-empowered interactive support on university EFL students' self-regulated strategy use and engagement in reading. The Internet and Higher Education, 65(1), Article 100991. https://doi.org/10.1016/j.iheduc.2024.100991

69.

* Perikos

Grivokostopoulou

Hatzilygeroudis

(2017). Assistance and feedback mechanism in an intelligent tutoring system for teaching conversion of natural language into logic. International Journal of Artificial Intelligence in Education, 27(3), 475–514. https://doi.org/10.1007/s40593-017-0139-y

70.

Pintrich

P. R.

(1995). Understanding self-regulated learning. New Directions for Teaching and Learning, 1995(63), 3–12. https://doi.org/10.1002/tl.37219956304

71.

Pintrich

P. R.

(2000). The role of goal orientation in self-regulated learning. In Boekaerts

Pintrich

P. R.

Zeidner

(Eds.), Handbook of self-regulation (pp. 451–502). Academic Press.

72.

Prasse

Webb

Deschênes

Parent

Aeschlimann

Goda

Yamada

Raynault

(2024). Challenges in promoting self-regulated learning in technology supported learning environments: An umbrella review of systematic reviews and meta-analyses. Technology, Knowledge and Learning, 29(4), 1809–1830. https://doi.org/10.1007/s10758-024-09772-z

73.

Russell

S. J.

Norvig

(2021). Artificial intelligence: A modern approach (4th ed.). Pearson.

74.

Sardi

Eliza

Darmansyah Candra

Yuliana

D. F.

Habibullah Yanto

D. T. P.

Eliza

(2025). How generative AI influences students’ self-regulated learning and critical thinking skills? A systematic review. International Journal of Engineering Pedagogy, 15(1), 94–108. https://doi.org/10.3991/ijep.v15i1.53379

75.

* Serrano

M.‐Á.

Vidal‐Abarca

Ferrer

(2018). Teaching self‐regulation strategies via an intelligent tutoring system (TuinLECweb): Effects for low‐skilled comprehenders. Journal of Computer Assisted Learning, 34(5), 515–525. https://doi.org/10.1111/jcal.12256

76.

* Shafiee Rad

(2025). Reinforcing L2 reading comprehension through artificial intelligence intervention: Refining engagement to foster self-regulated learning. Smart Learning Environments, 12(1), 23. https://doi.org/10.1186/s40561-025-00377-2

77.

* Su

J.-M.

(2020). A rule-based self-regulated learning assistance scheme to facilitate personalized learning with adaptive scaffoldings: A case study for learning computer software. Computer Applications in Engineering Education, 28(3), 536–555. https://doi.org/10.1002/cae.22222

78.

* Sun

Zhang

Liu

Zhang

(2025). How self-regulated learning is affected by feedback based on large language models: Data-driven sustainable development in computer programming learning. Electronics, 14(1), 194. https://doi.org/10.3390/electronics14010194

79.

Tomczak

J. M.

(2022). Deep generative modeling (1st ed.). Springer. https://doi.org/10.1007/978-3-030-93158-2

80.

* Tsai

C.-W.

Lee

Lin

M. Y.-C.

Cheng

Y.-P.

Lin

C.- H.

Tsai

M.-C.

(2025). Effects of integrating self-regulation scaffolding supported by chatbot and online collaborative reflection on students’ learning in an artificial intelligence course. Computers & Education, 232(1), Article 105305. https://doi.org/10.1016/j.compedu.2025.105305

81.

United Nations General Assembly . (2015). Transforming our world: The 2030 Agenda for Sustainable Development (A/RES/70/1). https://docs.un.org/en/A/RES/70/1

82.

* Vertemati

Zuccotti

Porrini

(2024). Enhancing anatomy education through flipped classroom and adaptive learning. A pilot project on liver anatomy. Journal of Medical Education and Curricular Development, 11(1), 1–7. https://doi.org/10.1177/23821205241248023

83.

Wang

Huang “Tammy

Sommer

Pei

Shidfar

Rehman

M. S.

Ritzhaupt

A. D.

Martin

(2024). The efficacy of artificial intelligence-enabled adaptive learning systems from 2010 to 2022 on learner outcomes: A meta-analysis. Journal of Educational Computing Research, 62(6), 1568–1603. https://doi.org/10.1177/07356331241240459

84.

* Wang

W.-S.

Lin

C.-J.

Lee

H.-Y.

Hwang

Y.-M.

T.-T.

(2025). Enhancing self-regulated learning and higher-order thinking skills in virtual reality: The impact of ChatGPT-integrated feedback aids. Education and Information Technologies, 30(14), 19419–19445. https://doi.org/10.1007/s10639-025-13557-x

85.

* Wang

Zheng

Tan

Lajoie

S. P.

(2023). Computer‐based scaffoldings influence students’ metacognitive monitoring and problem‐solving efficiency in an intelligent tutoring system. Journal of Computer Assisted Learning, 39(5), 1652–1665. https://doi.org/10.1111/jcal.12824

86.

* Wang

Zuo

Zou

Liu

Zhou

Zheng

Zhang

Yuan

Ren

(2024). Enhancing self-directed learning with custom GPT AI facilitation among medical students: A randomized controlled trial. Medical Teacher, 47(7), 1126–1133. https://doi.org/10.1080/0142159X.2024.2413023

87.

* Weber

Wambsganß

Söllner

(2025). Enhancing legal writing skills: The impact of formative feedback in a hybrid intelligence learning environment. British Journal of Educational Technology, 56(2), 650–677. https://doi.org/10.1111/bjet.13529

88.

Winne

P. H.

(2017). Learning analytics for self-regulated learning. In Siemens

Lang

(Eds.), Handbook of learning analytics (pp. 241–249). Beaumont, AB: Society for Learning Analytics Research.

89.

Winne

P. H.

(2018). Cognition and metacognition within self-regulated learning. In Schunk

D. H.

Greene

J. A.

(Eds.), Handbook of self-regulation of learning and performance (2nd ed., pp. 36–48). Routledge/Taylor & Francis Group. https://doi.org/10.4324/9781315697048-3

90.

Winne

P. H.

Hadwin

A. F.

(1998). Studying as self-regulated learning. In Hacker

D. J.

Dunlosky

Graesser

A. C.

(Eds.), Metacognition in educational theory and practice (pp. 277–304). Lawrence Erlbaum Associates Publishers.

91.

* Wu

T.-T.

Lee

H.-Y.

P.-H.

Huang

C.-N.

Huang

Y.-M.

(2024). Promoting self-regulation progress and knowledge construction in blended learning via ChatGPT-Based learning aid. Journal of Educational Computing Research, 61(8), 3–31. https://doi.org/10.1177/07356331231191125

92.

Xia

Liu

Tlili

Chiu

T. K. F.

(2026). A systematic literature review on designing self-regulated learning using generative artificial intelligence and its future research directions. Computers & Education, 240(1), Article 105465. https://doi.org/10.1016/j.compedu.2025.105465

93.

* Xu

Zhu

Wang

Qian

Yang

Zhang

(2024). The impact of a digital game-based AI chatbot on students’ academic performance, higher-order thinking, and behavioral patterns in an information technology curriculum. Applied Sciences, 14(15), 6418. https://doi.org/10.3390/app14156418

94.

Yang

Some

Bain

Kang

(2025). A comprehensive survey on integrating large language models with knowledge-based methods. Knowledge-Based Systems, 318(1), Article 113503. https://doi.org/10.1016/j.knosys.2025.113503

95.

* Yan

Lin

Kinshuk (2024). Adaptive Practicing Design to Facilitate Self-Regulated Learning Conception d'une pratique adaptative pour faciliter l’apprentissage autorégulé. Canadian Journal of Learning and Technology, 50(3), 1–22. https://doi.org/10.21432/cjlt28768

96.

* Yin

Pan

(2025). Effects of different AI-driven Chatbot feedback on learning outcomes and brain activity. Npj Science of Learning, 10(1), 17. https://doi.org/10.1038/s41539-025-00311-8

97.

* Zhang

Hew

K. F.

(2025). Leveraging unlabeled data: Fostering self-regulated learning in online education with semi-supervised recommender systems. Education and Information Technologies, 30(6), 7117–7142. https://doi.org/10.1007/s10639-024-13111-1

98.

* Zhu

(2025). Leveraging ChatGPT to support self-regulated learning in online courses. TechTrends, 69(5), 914–924. https://doi.org/10.1007/s11528-025-01075-z

99.

* Zhu

Wang

Bao

(2025). Using AI chatbots in visual programming: Effect on programming self-efficacy of upper primary school learners. International Journal of Information and Education Technology, 15(1), 30–38. https://doi.org/10.18178/ijiet.2025.15.1.2215

100.

Zimmerman

B. J.

(2000). Attainment of self-regulation: A social cognitive perspective. In Boekaerts

Pintrich

P. R.

Zeidner

(Eds.), Handbook of self-regulation (pp. 13–39). Academic Press.

101.

Zimmerman

B. J.

(2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70. https://doi.org/10.1207/s15430421tip4102_2

Components/Stages	Stage 1: Task definition	Stage 2: Goals & plans	Stage 3: Enactment	Stage 4: Adaptation
Conditions	Source of Knowledge (A = 1 \| G = 4) ^(S)	Source of Knowledge (A = 2 \| G = 3) ^(S)	Source of Knowledge (A = 23 \| G = 33) ^(S)	Source of Knowledge (A = 3 \| G = 4) ^(S)
Conditions	Source of Instructions (A = 2 \| G = 1) ^(S)	Source of Instructions (A = 6 \| G = 6) ^(S)	Source of Instructions (A = 34 \| G = 22) ^(S)	Source of Instructions (A = 5 \| G = 4) ^(S)
Operations	Cognitive Scaffolding (A = 3 \| G = 4) ^(S) Q&A and Learning Content Generation (G = 1) ^(G)	Cognitive Scaffolding (A = 1) ^(A) Learning Management (A = 2 \| G = 2) ^(S) Metacognitive Scaffolding (A = 5 \| G = 4) ^(S)	Affective Scaffolding (A = 1 \| G = 1) ^(S) Cognitive Scaffolding (A = 9 \| G = 13) ^(S) Learning Management Support (A = 2) ^(A) Metacognitive Scaffolding (A = 15 \| G = 5) ^(S) Negotiated Help-seeking Regulation (A = 1) ^(A) Q&A and Learning Content Generation (A = 3 \| G = 18) ^(S) Teachable Agent (A = 1) ^(A) Test & Exercise Generation (A = 5 \| G = 6) ^(S)	Affective Scaffolding (A = 2) ^(A) Metacognitive Scaffolding (A = 4 \| G = 2) ^(S) Test & Exercise Generation (G = 1) ^(G) Automated Metacognitive Report (A = 3 \| G = 3) ^(S)
Products		Learning Path Products (A = 1 \| G = 2) ^(S)	Assignment Product Generation (G = 2) ^(G)	Learning Path Products (A = 1 \| G = 1) ^(S)
Evaluations			Affective feedback (A = 2 \| G = 2) ^(S)
			Cognitive Feedback (A = 14 \| G = 13) ^(S)
			Grading (A = 6 \| G = 1) ^(S)
			Metacognitive Feedback (A = 15 \| G = 2) ^(S)
Standards		Task Prioritization (A = 1) ^(A)
Standards		Benchmark Setting (A = 2) ^(A)