Generative AI supports self-regulated learning across three phases—forethought, performance, and self-reflection—through six pedagogical affordances, with the most popular student activities being information searching in forethought, problem-solving strategies in performance, and feedback/self-assessment in self-reflection, though effectiveness depends critically on individual factors (self-efficacy, autonomy, competence) and environmental support (technical quality, teacher guidance, peer collaboration).
Objective: This systematic literature review investigated how to design self-regulated learning (SRL) activities using generative AI (GenAI). Despite growing evidence that GenAI can enhance SRL through immediacy and interactivity, critical gaps remain: mechanisms by which GenAI influences SRL are unclear, and teachers struggle to incorporate these tools effectively into classrooms. The study aimed to identify the pedagogical affordances of GenAI across SRL phases, catalog learning activities GenAI fosters, determine factors influencing SRL effectiveness with GenAI, and provide practical guidance for educators and tool developers.
Methods: The review followed PRISMA guidelines with three major steps: literature search, screening, and coding. Four databases were searched—Web of Science, ProQuest, ERIC, and Scopus—using a comprehensive query combining situational variables (SRL terms like self-regulation, metacognition, goal-setting), intervention variables (AI terms like ChatGPT, generative AI, large language models), and individual variables (education terms like student, learning, teaching).
The search was conducted July 11, 2025, covering publications from January 1, 2020 to December 31, 2024. This timeframe was chosen because most relevant research occurred after ChatGPT's launch in November 2022, making the past three years particularly rich in GenAI applications to education. The initial retrieval yielded 8,684 articles: 2,976 from Web of Science Core Collection, 4,923 from Scopus, 466 from ProQuest Dissertations & Theses Global, and 319 from ERIC.
Selection involved three steps. Initial retrieval removed duplicates (n=769), articles outside the date range (n=3,378), non-journal articles (n=1,736), and non-English publications, leaving 2,261 for abstract screening. Abstract screening excluded articles for: not written in English (n=5), unrelated to GenAI (n=203), unrelated to SRL (n=1,763), and non-empirical design (n=144), yielding 146 for full-text assessment. Full-text assessment removed: unavailable texts (n=6), non-English (n=5), unrelated to GenAI (n=22), unrelated to SRL (n=25), non-empirical (n=12), and duplicate texts (n=3), resulting in 73 articles for final analysis.
Two coders independently analyzed all articles using established frameworks. For SRL phases, they applied Zimmerman and Moylan's (2009) social cognitive model with three phases: forethought (coded as "1"), performance ("2"), and self-reflection ("3"). For learning activities, they adopted Chiu's (2024) Delphi study classification suggesting 20 learning activities with ChatGPT: four related to forethought, eight to performance, and eight to self-reflection, coded 1-20.
Pedagogical affordances were derived inductively through bottom-up coding, identifying six major affordances coded 1-6. Learning outcomes were divided into cognitive, affective, and behavioral dimensions using Bloom's taxonomy. The Mixed Methods Appraisal Tool (MMAT) evaluated bias risk for each study. Any coding discrepancies between the two coders were resolved through discussion until consensus was reached.
Key Findings:
Descriptive Statistics: The 73 reviewed articles were published across 46 different journals, predominantly after 2023 (68 articles). Most used chatbots like ChatGPT, though other GenAI applications included intelligent tutoring systems, AI coaches/virtual humans, and AI visual feedback. The disciplines spanned languages, sports, computer science, and mathematics. Twelve articles focused on K-12 education while 58 focused on higher education.
Research methodology varied: 36 studies used intervention designs, 14 used questionnaires only, and 12 used mixed methods. Sample sizes ranged dramatically from 6 to 1,687 participants, with most exceeding 30. Students were predominantly from Australia, though some came from New Zealand and the UK. The wide range of contexts and methodologies provided a comprehensive view of GenAI applications in SRL.
RQ1: Six Pedagogical Affordances Across SRL Phases (see Figure 2 in original): Analysis revealed GenAI provides six major pedagogical affordances distributed unevenly across the three SRL phases:
(i) Creating personalized learning objectives (n=6, least frequent): GenAI helps students set clear, personalized, and realistic goals. Traditional students with low self-regulation often set unrealistic or overly ambitious goals. For example, Mohan et al. (2020) developed an adaptive goal-setting method using an intelligent health coach that interacted continuously with participants, setting initial goals, dynamically adjusting them, and presenting goals interactively. The coach generated exercise goals tailored to individual participants that aligned with clinical expert recommendations. Similarly, Hew et al. (2023) developed a goal-setting chatbot grounded in SMART principles, guiding users through five structured questions with predefined options. By collecting data about students' learning preferences and knowledge levels through human-computer interaction, the chatbot provided goal suggestions suited to each student's needs.
(ii) Search, analysis, and integrate resources (n=17): GenAI offers distinct advantages over traditional keyword-based search methods. Traditional approaches often miss semantically related content lacking direct keyword matches. In contrast, GenAI's advanced pattern recognition enables semantic-level searches delivering relevant results based on user intent. Durall Gazulla et al. (2023) demonstrated students could use a chatbot to find information on goal-setting, task organization, and overcoming obstacles, surpassing basic keyword searches. GenAI provides real-time, convergent information to learners' inquiries, minimizing interruptions during the performance phase. In data analysis, GenAI leverages multi-layer neural networks to mine deeper features and abstract concepts while uncovering potential causal relationships across diverse data types (text, images, audio). Regarding information integration, GenAI creatively combines data to generate novel insights. Pan et al. (2024) showed the ReadMate platform adapted reading materials to students' levels and preferences, demonstrating GenAI's ability to search, analyze, integrate, and generate valuable, context-specific information.
(iii) Monitoring and evaluating progress (n=26, most frequent): This affordance plays important roles in all three SRL stages, reflecting GenAI's capacity to oversee the entire learning process. GenAI provides reports and evaluations of students' behavior during each learning cycle. When students face difficulties executing plans, GenAI detects problems and suggests alternative strategies, helping students adjust and strengthen plans to achieve goals more effectively. Many researchers combine dashboard visualization with AI to monitor students' learning processes. Some GenAI tools like empathic chatbots can evaluate students' emotional and cognitive states by analyzing language and behavior patterns beyond behavioral data alone. This emotional information assists AI in adjusting teaching methods or providing encouragement, better supporting students' self-regulation.
(iv) Recommending learning strategies (n=11): This has received less attention from researchers. Learning strategies are influenced by individual preferences and learning styles. Students often treat GenAI as an "answer generator," asking for solutions or problem-solving processes, rarely engaging with AI as a "strategy advisor" that teaches them how to learn. This reflects two main issues: AI tools have unclear positioning and insufficient popularity in education, leading to lack of awareness about AI's potential for enhancing metacognitive abilities; and students' questioning capabilities limit AI's learning strategy recommendation effectiveness, often producing vague outputs or failing to consider psychological states. Poor experiences decrease students' willingness to ask for learning strategies. Despite limited attention, Liu, Zhang, and Biebricher (2024) found large models can promote students' use of SRL strategies, improving the learning process—highlighting potential for AI to guide students toward developing and refining learning strategies.
(v) Recording progress and providing feedback (n=25): Similar to monitoring, researchers focus on this during self-reflection because this phase serves as the final step in the cyclical process requiring feedback loop formation. Through ChatGPT feedback, students enhance language skills, self-efficacy, and metacognition. Banihashem et al. (2024) revealed significant differences between peer feedback quality and AI-generated feedback: GenAI tends to produce more general descriptive feedback including summary statements, while students perform better in precisely identifying and pinpointing issues. Liu et al. (2021) proposed a reflective thinking promotion mechanism-based AI-supported English writing (RTP-AIEW) approach. Compared to conventional AI-supported writing, RTP-AIEW improved English writing performance, self-efficacy, and SRL while significantly reducing cognitive load. Liao et al. (2024) and Sun et al. (2023) utilized AI visualization as formative assessment, helping students understand current learning levels and cognitive states, prompting examination and reflection on performance. Afzaal et al. (2024) designed an algorithm identifying root causes of declining performance, automatically generating data-driven action recommendations. Some studies set up functional reflection windows in learning platforms where students independently feedback learning experiences to AI, which then recommends new learning goals and action plans for students to accept or reject based on independent judgment.
(vi) Generating new ideas and examples (n=16): By analyzing students' work and offering insights or alternative perspectives, GenAI tools prompt students to consider different approaches or solutions, stimulating creativity and encouraging exploration of concepts they might not have thought of independently. Additionally, AI supports metacognition by guiding students to reflect on thought processes, leading to deeper understanding and development of new viewpoints. Fırat and Kuleli (2024) compared ChatGPT and Google Search for undergraduate students' SRL of JavaScript functions, finding ChatGPT provided tailored insights and correct code examples, reducing cognitive load. Nguyen and Nguyen (2024) studied how 17 students and professionals used GenAI for brainstorming and reflective design practices, finding collaboration with AI tools enhanced decision-making and refined design ideas. These tools enhance organizations' ability to integrate external knowledge, allowing employees to continuously update skills and knowledge to adapt to market changes, driving innovation more effectively.
Overall, GenAI tools often play multiple roles in SRL, positively impacting students' performance by helping set goals, assisting in planning, monitoring actions, and reflecting on progress. This highlights the importance of guiding students to understand and use these diverse pedagogical affordances.
RQ2: Twenty Learning Activities Across SRL Phases (see Figure 3 and Table 4): Among reviewed articles, 30 conceptualize SRL as a whole concept without specifying phases; 19 incorporate GenAI into all three phases; others address one or two phases. The distribution shows distinct activity patterns per phase.
Forethought phase (27 articles): Two main areas—task analysis and self-motivation beliefs. In task analysis, GenAI served as AI coaches to: (i) help users set learning goals and choose strategies through dialogue; (ii) analyze task objectives, requirements, difficulty, search available resources, and assess challenges; (iii) encourage personalized learning to use strategies and achieve goals. Students primarily engaged in searching for information (n=26, most popular) and creating lists/outlines (n=7). Other activities included getting examples, asking for definitions, and generating discussion questions.
In self-motivational beliefs, 11 articles suggested GenAI enhanced students' motivation and self-efficacy during forethought by answering questions and giving feedback. GenAI helps students believe in their capacity to achieve specific goals.
Performance phase (31 articles): Students used GenAI to enhance learning through various self-regulation strategies based on self-observation and self-control. This phase showed the most diverse activity range with eight different types observed:
(i) Time management: Takagi et al. (2023) indicated students managed time through study plans and learning records provided by GenAI. GenAI-based chatbots helped manage time, monitor progress, regulate emotions, focusing on stress management and maintaining interest/motivation.
(ii) Environmental structuring: Includes learning resources, learning environment, and external feedback. As an external learning environment, GenAI provides materials, resources, and feedback. Hu et al. (2024) showed chatbots establish social connections with students, significantly improving affective states and learning efficiency.
(iii) Help-seeking: Twelve studies required students to seek GenAI help completing tasks/goals in writing, reading, and research scenarios. Getting insight into complex problems (n=8) was popular. Chen et al. (2025) demonstrated students exhibit pragmatic help-seeking behaviors when interacting with large models, asking operational questions while bypassing metacognitive phases outlined in traditional help-seeking models (diagnosing problems, evaluating assistance), revealing non-linear features previous models failed to capture. Neumann et al. (2025) integrated LLMs into learning management systems supporting SRL and help-seeking, achieving 88% accuracy in providing course-related assistance.
The most popular performance activity was using strategies for problem-solving, particularly searching information (n=26), getting insight into problems, and asking for ideas for improvement (n=19).
Self-reflection phase (35 articles): Two main areas—self-judgment and self-reaction.
In self-judgment, GenAI records and analyzes students' behavioral data, assignments, or test results through data charts, evaluation reports, and progress bars. By comparing these with preset goals, students assess their own performance and learning progress.
In self-reaction, students respond based on current situation. By answering review questions, practicing new problems, and receiving GenAI feedback, students decide whether to adopt, modify, or reject feedback based on own judgment. Self-reaction includes emotional responses and behavioral adjustments. Currently, GenAI tools play guiding roles by recommending learning paths, supplementing resources, and providing emotional support, guiding students to make positive reactions based on feedback.
The main self-reflection activities were getting feedback for their work (n=28, most popular across all phases) and asking for ideas for their improvement (n=19). Other activities included checking answers (n=10), summarizing own work (n=9), and generating review questions (n=6).
RQ3: Individual and Environmental Factors Influencing Effectiveness: Focusing on the 62 quantitative studies (84.93%), researchers extracted and categorized 55 potential influencing factors into two dimensions.
Individual dimensions: Typical factors included self-efficacy (n=7, most frequent individual factor), perceived autonomy (n=4), perceived competence (n=3), perceived relatedness (n=3), interest (n=3), intrinsic motivation (n=2), and strategy use (n=2).
(i) Self-Efficacy: Ghaleb and Alshiha (2023) and Wu et al. (2024) studied self-efficacy as a mediator between AI and SRL, finding it significantly impacts SRL ability.
(ii) Needs satisfaction: Autonomy, competence, and relatedness are basic psychological needs. Zhou and Zhang (2024) found strong connections between perceived autonomy and intrinsic motivation, both being potential SRL effectiveness factors. Gender is also significant. Xia, Chiu, Chai, & Xie (2023) found autonomy and competence needs fulfillment is moderated by gender and AI knowledge, while relatedness needs fulfillment is only moderated by gender. Among girls with insufficient AI knowledge, autonomy and competence more strongly predict SRL. Teachers can design activities supporting students' autonomy, competence, and relatedness to enhance learning motivation.
(iii) Interest: Li, Sadiq, Qambar, & Zheng (2025) explored intrinsic motivation's mediating role between ChatGPT usage and SRL, finding motivation mediated the relationship, both having positive impact. Motivation influences how students set goals, select strategies, and show resilience overcoming obstacles.
(iv) Learning strategy capacity: An essential SRL feature is continuous strategy adjustment and refinement based on feedback. Huang et al. (2023) discussed four learning strategies' (repetition, elaboration, organization, critical thinking) impact on review activities, finding the interaction between strategies and review activities significantly influenced academic performance.
Environmental dimensions: Peers (n=4) and AI (n=6) were major foci. Three key aspects emerged: technical support, teacher support, and peer support.
(i) Technical Support: Technology acceptance is influenced by perceived ease of use and perceived usefulness. Platform environment supporting GenAI is an influencing factor, including platform quality and ease of use. Zhou and Zhang (2024) highlighted perceived usefulness plays significant roles in encouraging continuous intention to use technology. Martins et al. (2024) indicated 35% of conversations ended before users actually interacted with chatbots, suggesting students might click on new tools out of curiosity yet discontinue due to lack of familiarity or interest. Glick et al. (2024) found students unfamiliar with GenAI platform tools who haven't completed technical training have affected participation and engagement in planning and self-reflection stages.
Many studies pointed out GenAI might enhance over-reliance, leading to "cognitive laziness" potentially hindering self-regulation and deep learning abilities. Although numerous studies demonstrated significant short-term benefits on learning performance, long-term effects remain concerning. Tailored or self-developed large language models have stronger impacts promoting SRL than general-purpose models. Sun et al. (2023) used a personalized learning system with three components: prior knowledge assessment, visualization features (performance comparison with peers and previous self, assessment summaries), and self-regulation strategies (goal setting, time management, countdown timer, note-taking, unit selection). Future research could focus more on tailored or self-developed models' roles facilitating SRL.
(ii) Teacher Support: Only few studies discuss teachers' roles, mainly providing: timely feedback, emotional support, and learning strategy guidance. Teachers need to provide appropriate technology guidance and training. In included articles' instructional designs, teachers played active roles offering guiding materials, helping students use GenAI tools, assisting in discussions and demonstrations, engaging in collaborative learning and reflection alongside students. Wu, Zhang, et al. (2024) indicated teacher support significantly and positively influences learning strategies and motivation, directly and indirectly influencing autonomous learning ability. Jin et al. (2023) learned from interviews that students prefer teacher guidance and hope AI provides content matching teacher-given test levels.
(iii) Peer Support: Younis (2024) explored self-regulation level differences among students learning independently, with peers, and in groups under ChatGPT support, finding peer and group learning more effective fostering self-regulation skills than independent learning. Su et al. (2024) researched cooperative games versus direct instruction impacts on kindergarten children's computational thinking, sequencing, self-regulation, and theory of mind in early AI education, showing direct instruction more effective than cooperative play improving computational thinking and self-regulation. Wu, Zhang, et al. (2024) noted teacher support serves as mediating variable between AI and SRL.
Implications: The findings help understand how AI as a human-machine collaborative tool affects SRL processes, providing specific pedagogical affordances, learning activities, and influential factors directly affecting student SRL.
Four Practical Suggestions: First, teacher professional learning programs should discuss pedagogical affordances of GenAI so teachers better guide students in SRL using GenAI—for example, exploring how ChatGPT assists self-reflection or how Pictory supports multimodal project planning. Second, teachers should give specific guidance on how students interact with GenAI in each stage: using GenAI to set goals or co-create study plans in forethought; receiving real-time support like content summaries during performance; and prompting reflective journaling to analyze learning processes in self-reflection. Third, teachers should utilize both individual and environmental dimensions to establish engaging SRL environments—integrating students' individual goals and psychological needs (e.g., autonomy) with supportive classroom environments encouraging collaboration and providing GenAI tools. Finally, the three suggestions work best together—teachers should understand how they relate and plan curriculum accordingly.
Visualization of Findings: Figure 4 shows how the three empirical findings relate: GenAI's pedagogical affordances support student learning activities for SRL. Students may engage or disengage in activities, influenced by individual (self-efficacy, autonomy, competence, relatedness, interest, intrinsic motivation, strategy use) and environmental dimensions (technical support, teacher support, peer support). The SRL cycle still requires students to set goals, assess progress, and adjust strategies, with GenAI providing diverse choices and personalized suggestions. Students remain decision-makers choosing appropriate goals and strategies. The entire SRL process is still dominated by students across all three phases: in forethought, students autonomously evaluate and select information generated by GenAI; in performance, students remain subjects of actual learning activities while GenAI helps answer questions, provide resources, and monitor; in self-reflection, GenAI facilitates self-assessment through automatic feedback, reflective scaffolding, or personalized scaffolding, with students deciding whether to adopt and how to adjust subsequent learning.
Limitations: The study acknowledges two main constraints. First, the literature search was limited by search keywords and completed in October 2024, so newer articles were not included. New GenAI developments like large language models, image generation, and video generation may add more pedagogical affordances and learning activities. Second, the review analyzed articles from educational perspectives, but engineering perspectives play important roles in SRL that weren't fully explored.
Future Directions: The researchers suggest five major future research directions:
More research on pedagogical affordances of GenAI: This review offers six major affordances, mostly related to large language model chatbots. The suggested affordances may be driven only by chatbots. Future research should include more GenAI tools such as text-to-images, text-to-videos, and specific GenAI tools for coding, languages, and mathematics.
More research on student SRL activities using GenAI: This review used 20 learning activities to analyze articles. Future research should explore more student learning activities for each SRL phase.
More research on environmental dimensions: Most studies concerned individual dimensions. More studies needed on how to design environments to better engage students in SRL using GenAI—for example, roles of teachers and peers, policy, and assessment.
More research on relationships among the three empirical findings: The relationships among components in Figure 4 are not investigated. Future studies should use correlational designs to examine relationships.
More research on dynamic interplay between GenAI features and SRL phase-specific needs: Studies could use diverse research methods like Ordered Network Analysis or combinations of storyboards and speed dating methods to explore how GenAI adaptability (e.g., real-time feedback, resource recommendation) aligns with distinct needs of Zimmerman's three SRL phases. This would clarify whether phase-specific tailored GenAI interventions can more effectively promote SRL development than generic support.
Title and Authors: "A systematic literature review on designing self-regulated learning using generative artificial intelligence and its future research directions" by Qi Xia and Qian Liu (Department of Higher Education, Zhejiang University, Room 506-2, School of Education, Hangzhou 310058, China), Ahmed Tlili (Smart Learning Institute, Beijing Normal University, 12th Floor, Tower A, Jingshi Technology Building, 12 Xueyuan South Road, Haidian District, Beijing, 100082, China), and Thomas K.F. Chiu (Department of Curriculum and Instruction, Centre for Learning Sciences and Technologies, Centre for University and School Partnership, The Chinese University of Hong Kong, Hong Kong SAR) (corresponding author).
Published On: Received February 25, 2025; Received in revised form September 21, 2025; Accepted September 22, 2025; Available online September 22, 2025.
Published By: Computers & Education Volume 240 (2026) 105465, published by Elsevier Ltd. DOI: https://doi.org/10.1016/j.compedu.2025.105465. All rights are reserved, including those for text and data mining, AI training, and similar technologies. The research was supported by the Zhejiang Province Soft Science General Project (Grant No. 2025C35084), titled "Exploring the Mechanisms of Generative Artificial Intelligence in Empowering Interdisciplinary Education: A Strategic Approach to Strengthening Higher Education in Zhejiang Province." No human participants were involved, and datasets used are available from the corresponding author upon reasonable request.