Blended learning plays an important role in the digital transformation of higher education and requires students to regulate their learning with greater independence. Although learning analytics studies have increasingly used clickstream data to identify self-regulated learning (SRL) tactics and strategies, many rely on platform-specific behavioral categories whose theoretical meaning is not always explicit. Moreover, less is known about how SRL strategies change across the structured phases of compulsory courses. This study examined the online behaviors of first-year undergraduates in an institution-initiated blended general education course at a Chinese university. Drawing on Pintrich's SRL framework, five learning tactics were identified and consolidated into four strategies, by using sequence analysis and clustering. The findings indicate that students frequently relied on strategies characterized by limited regulation and oriented toward rapid task completion. Strategy use also followed distinct phase-specific trajectories, that is many students began with cognitively demanding approaches, shifted mid-course to lower-order strategies, and then returned to deeper forms of engagement during exam preparation. Although a considerable proportion of students shifted toward strategies characterized by lower levels of regulation, upward shifts toward elaboration-oriented regulation were associated with better academic performance. These findings suggest that SRL dynamics in blended learning are shaped by both learners' regulatory choices and the temporal structure of course design, highlighting the need for instructional scaffolding that supports the sustained use of higher-order strategies.
The increasing adoption of online teaching in higher education (HE), especially in the post-pandemic era and with the rise of artificial intelligence, highlights the need to better understand how online teaching expertise evolves. Recognising that such developments may fluctuate across contexts and time, this study employed an experience sampling method (ESM) to investigate both within-teacher variability and between-teacher differences in time-variant predictors of perceived expertise development in online teaching. The sample consisted of 25 university teachers, representing a substantial size for intensive longitudinal research using ESM in HE. Participants provided repeated diary entries via a mobile app before and after teaching or planning sessions. These entries captured emotional valence, digital skills, reflection and perceived development of online teaching expertise. Bayesian cumulative mixed-effects analyses revealed that state positive emotions during teaching and planning sessions were positively associated with perceived development of online teaching expertise. Momentary perceptions of digital skilfulness during sessions were more influential than stable trait-level differences between individuals. Furthermore, teachers who engaged more frequently in reflective practices across sessions reported greater cumulative expertise development. These findings underline the importance of withinteacher dynamics in emotions and digital skills and between-teacher differences in reflection in shaping the development of online teaching expertise. The implications of the study and future research are discussed.
Conversational agents are shifting from solo tutors to teammates in online project-based learning, yet little is known about how design features shape dynamics. Drawing on self-regulated learning and volitional control theory, we examine how ChatGPT affects team learning. A survey of 356 students nested in 48 teams tests a model linking AI explainability to intention to continue learning via perceived social presence, with team perceived virtuality and cohesion as moderators. Explainability heightens the agent's social presence, and this copresence boosts intention to continue learning. High virtuality and cohesion strengthen this pathway. Social presence is pivotal, and we offer design recommendations for embedding explainable AI in teamwork.
This study investigates how specific prompting behaviors influence the academic performance of engineering students using ChatGPT. A total of 128 fourth-year students across four engineering programs participated in weekly sessions involving rotating task types: case analysis, engineering design, multi-step problem-solving, and experimental data analysis. A Python-based interface connected students to ChatGPT-4o via the API, logging all interactions and implementing stratified randomization of AI access across the 16-week semester. Written assignments were used for both grading and analyzing AI content integration. Ten metrics were designed to capture student behaviors—eight focusing on prompting (AI Query Count, Query Depth and Structure, Query Efficiency, Prompt Refinement Depth, Response Utility, Response Complexity, Response Reliance, and AI-Driven Problem-Solving) and two assessing writing quality (Structural Complexity Score and Content Richness and Information Density Score). Inter-metric analysis examined how these behaviors related to one another and to academic performance. Results showed that AI Query Efficiency, which reflects how effectively students crafted clear, well-structured prompts, and AI-Driven Problem-Solving, which captures how strategically they integrated AI output into their reasoning, were the strongest predictors of academic success. These findings were supported by Mann–Whitney U tests comparing AI and non-AI groups, as well as Spearman correlations, Random Forest regressors, partial dependence plots, principal component analysis, and mixed-effects modeling, and remained significant even after accounting for students' cumulative GPA. The results suggest that prompting strategy plays a meaningful role in shaping how effectively students use AI in engineering education.
As generative artificial intelligence (AI) tools become increasingly embedded in higher education, concerns are growing over students' psychological dependency on these technologies. Drawing on stress-coping theory, this study examines how academic expectation stress contributes to generative AI dependency among university students. Using survey data from 1,531 students across ten Chinese universities, three mediating mechanisms linking stress to dependency has been identified: increased frequency of generative AI use, heightened reliance intention, and enhanced knowledge of generative AI's limitations. While the first two intensify dependency, the third mitigates it. Notably, study engagement-traits typically considered protective-paradoxically amplify the stress-to-dependency relationship. These findings advance understanding of student-AI relationships and highlight how high-pressure educational environments may unintentionally reinforce overreliance on generative AI tools.
Few studies have investigated how digital inequalities and hybrid learning demands jointly shape student wellbeing in rural higher education, particularly in emerging economies where structural barriers are most acute. This study addresses this gap by integrating the Conservation of Resources (COR) theory and the Job Demands-Resources (JD-R) model to explain the psychological mechanisms at play. An explanatory sequential mixed-method design was employed. In the quantitative phase, survey data from 688 undergraduate students across rural universities in Eastern Indonesia were analyzed using PLS-SEM. Results demonstrate that digital inequalities and hybrid learning demands significantly increase perceived stress, which in turn reduces student wellbeing. Mediation analysis confirmed stress as a key psychological pathway. In the qualitative phase, interviews and focus groups with 22 students revealed narratives of digital fatigue, disrupted academic routines, and varied coping strategies shaped by local context and gender. These findings not only enrich the quantitative evidence but also highlight the importance of institutional support in mitigating stress. By extending COR and JD-R frameworks with digital and contextual dimensions from the Global South, this study advances theoretical debates and provides actionable insights for universities and policymakers to design more inclusive and adaptive hybrid learning systems.
The role of the moderator is a key factor that influences the effectiveness of asynchronous online discussions (AODs), but its impact on different levels of discussion tasks remains unclear. To address this research gap, a 2 & times; 2 factorial experiment was conducted to explore the effects and interactions between the moderator (present vs. absent) and discussion task level (lower vs. higher) on learners' learning outcomes, experiences, participation behavior, and patterns of knowledge construction in AODs. A total of 109 undergraduate participants were randomly assigned to four conditions. The results indicated that neither the moderator nor the task level had any significant impact on learning outcomes. However, the moderator enhanced learners' affective experience and increased the total number of words posted, while the task level significantly influenced learners' initial posting, with higher-level tasks generating more posts. Epistemic network analysis (ENA) was employed to analyze participants' patterns of knowledge construction in AODs. The ENA results revealed that, in higher-level tasks, moderators effectively reduced the social coordination burden among learners, enabling them to redirect their cognitive focus toward meaning construction and higher-order inquiry. In contrast, in lower-level tasks, the moderator's presence appeared redundant. This study discusses how these findings can inform the design and execution of AODs.
Emotions are central to learning in higher education, influencing motivation, engagement, and achievement. As digital and hybrid instruction expand, however, students' affective cues often become obscured, limiting timely pedagogical support. Emotional Artificial Intelligence (E-AI) has emerged as a potential solution, capable of detecting and responding to emotional states through multimodal sensing and machine learning. Yet current research frequently conflates recognition accuracy with educational validity, treats facial expressions as universal indicators, underreports cultural bias, and advances in disciplinary silos. This study conducts a configurative systematic review of E-AI in higher education (2010-2025) across six major databases and grey literature, guided by PRISMA. Findings reveal three historical research waves and four dominant lenses: technical, pedagogical, theoretical, and ethical. While pedagogical applications dominate, theory integration and ethical safeguards lag behind. We propose a maturity framework positioning E-AI as a socio-technical system that must be theory-anchored, pedagogy-first, bias-aware, and ethically governed.
The rapid adoption of generative artificial intelligence (GenAI) in higher education environments is transforming how learners access guidance and feedback. Yet, students' AI self-efficacy (AISE) in using AI tools may critically influence how they engage with these technologies. This study developed a GenAI agent integrated with a structured knowledge graph (KG) to provide personalized, real-time feedback in an online Preschool Health and Hygiene course. Ninety-eight university students were grouped by AISE levels (high, medium, low) to explore differences in AI feedback adoption, critical thinking, and cognitive load during two learning tasks. The results indicated that high-AISE learners leveraged knowledge graph-integrated generative AI feedback more strategically, showing greater self-monitoring skills, improved self-evaluations, and lower cognitive load on complex tasks. It was also found that the high-AISE group showed higher self-evaluation ratings than the low-AISE group only on their second revised tasks, while no significant difference was found between the two groups in the first initial and revised tasks, or in the second initial task. Qualitative findings suggest evolving learner-agent relationships, shifting from reliance on AI authority toward reflective collaboration. In particular, the high-AISE group was more inclined to perceive GenAI as a collaborative learning partner, demonstrating higher-order abilities in information evaluation and integration. These insights highlight the need to scaffold AI literacy and tailor GenAI-supported learning environments to students' AISE levels, informing future designs of adaptive higher education systems. The findings also imply the importance of conducting further research to support students in improving their AI self-efficacy in the future.
Social annotation is a collaborative learning practice that engages students in active reading, in-text commenting, and peer dialogue around shared digital texts. While it holds promise for supporting deep discourse and learning, challenges persist, including inconsistent, superficial, and fragmented participation. This study investigates how analytics-based suggestions influence students' social annotation engagement (defined multi-dimensionally as their contributions, peer interaction, and reading activity) in a fully asynchronous undergraduate course. Using a within-subject phased design over nine weeks, trace and textual data were collected from 94 students to analyze the patterns of their engagement across three phases: pre-intervention, intervention, and post-intervention. Quantitative analyses, including linear mixed-effects modeling and social network analysis, revealed that analytics-based suggestions significantly enhanced student engagement during the intervention phase: students produced more frequent, longer, and higher-quality annotation contributions, participated more actively and densely in peer discussions, and spent more time reading. However, these improvements were not sustained after the intervention ended, pointing to a reliance on the suggestions rather than the development of lasting practices. Findings highlight the potential of analytics-informed support to catalyze social annotation learning while underscoring the challenges for sustaining this engagement over time. The study contributes both empirical evidence and design considerations for integrating learning analytics into social annotation learning contexts in ways that promote deeper, more consistent, and sustained engagement.
Student burnout in online learning is increasingly recognized as a barrier to engagement in learning and to academic success. Understanding how burnout manifests and develops in online learning contexts is essential for designing targeted interventions that support student wellbeing and performance. This study employed both graphical Gaussian models (GGMs) and directed acyclic graphs (DAGs) to explore the interplay among symptoms across three core dimensions of burnout: emotional exhaustion, improper behavior, and a low sense of achievement. Data were collected from 486 online learners through self-reported measures. Results showed that emotional exhaustion symptoms, particularly tiredness, boredom, and exhaustion, emerged as the most central and predictable nodes in the network. DAG analysis further revealed that symptoms from the emotional exhaustion dimension exerted the strongest directional influence on those in the improper behavior dimension. Together with existing theoretical frameworks, these findings highlight structures and sequential pathways of burnout dimensions, and provide actionable insights for designing early detection tools and targeted interventions in online higher education contexts.
Innovative technological tools such as virtual reality have the potential to transform educational practice. However, the complex ways in which teachers relate to innovative technologies remain under-investigated. This study examined how the technological pedagogical content knowledge (TPACK) and technology acceptance (TAM) of Finnish and Turkish preservice teachers are influenced by their participation in STEM activities in virtual reality environments. A convergent mixed-methods design was used, combining quantitative nonparametric tests conducted with TPACK and TAM scales and qualitative content analysis of participants' responses to open-ended questions. The findings revealed that interdisciplinary STEM activities conducted in virtual reality environments significantly improved the TPACK of both Finnish and Turkish preservice teachers. Changes in TAM showed different patterns between the countries and the qualitative findings provided possible contextual explanations for those patterns. The participating preservice teachers stated that virtual reality environments supported their skills in problem solving, conceptual understanding, technology use, and professional awareness, indicating that this type of technology can serve as a creative and effective educational tool.
Recent advances in generative AI (GenAI) technologies have reshaped student learning in higher education. However, most studies treat AI usage as a generalized construct, overlooking differences across various disciplines, educational levels, and regions. This gap is particularly relevant in China, where regional disparities in access and readiness may influence student engagement with AI tools. This study examines how university students in China perceive, utilize, and are influenced by generative AI tools across various disciplines and regions, while also assessing the accuracy and disciplinary relevance of AI-generated responses in professional coursework. The study includes four mixed-methods investigations: a national survey of students’ AI usage patterns and attitudes; an evaluation of ChatGPT’s responses to subject-specific questions based on semantic similarity and expert scoring; behavioral clustering and analysis of one-month AI usage logs to identify distinct user profiles; and an exploration of AI tools’ impact on academic performance and long-term outcomes. Findings show significant variation in AI engagement across disciplines and regions. Students in economically developed areas report higher usage and broader application, while those in central and western regions have more limited access. Students’ academic performance can be improved to varying degrees through AI-driven learning, with engineering and natural sciences students primarily using AI for technical tasks, while humanities and arts students employ it for linguistic and creative support. Evaluation of ChatGPT’s responses reveals moderate accuracy, with disciplinary variations. Despite generally positive attitudes, both students and experts express concerns about content reliability, over-reliance on AI, and its potential to hinder independent thinking and creativity.
Machine learning (ML) has become integral to online education, enhancing prediction, personalization, and automated assessment. However, algorithmic bias remains a critical barrier to equitable learning, as ML models can systematically under- or over-estimate outcomes for particular demographic groups. Existing fairness approaches—especially those focused on single attributes such as race or gender—fail to capture the complex, intersectional identities that shape students' experiences. Even multi-group fairness methods face key limitations in educational contexts, including computational scalability and difficulty adapting to shifting data distributions and fairness priorities. To address these challenges, this study proposes a reinforcement learning (RL)-based pre-processing framework that dynamically reweights data to optimize both predictive accuracy and multi-group fairness while safeguarding privacy. An AUC-based fairness metric ensures stability as subgroup combinations increase, and explainable AI (XAI) techniques enhance interpretability. Using large-scale data from Algebra Nation and state assessments, results show that the proposed framework achieves improved fairness and stable accuracy, offering a scalable, model-agnostic, and privacy-preserving pathway toward trustworthy AI in education.
Asynchronous online learning environments offer flexibility for students to navigate learning at their own pace, resulting in diverse behavioral patterns that can significantly impact cognitive load and academic performance. However, limited research has explored how learner-controlled environments shape these patterns and their relationship to learning outcomes and levels of cognitive load. This study investigates behavioral patterns in an online asynchronous graduate law class (n = 90) at a large university, analyzing data from a learning management system to categorize students into clusters based on their interactions with instructional components (e. g., video lectures, video-based examples, and problem-solving tasks). Cluster analysis revealed three distinct patterns: balanced learners, who achieved the highest performance; practice-oriented learners, who exhibited lower intrinsic cognitive load; and classic learners, characterized by comparatively lower extraneous cognitive load. However, these differences in extraneous load were not statistically significant between clusters, suggesting that the additional cognitive demands of learner control may have imposed similar baseline levels of extraneous load regardless of behavior pattern. Contrary to expectations, learner behavior patterns extended beyond example-based and problem-solving-first approaches, highlighting greater variability in learner strategies. These findings underscore the importance of understanding instructional paths in learner-controlled environments and how behavioral patterns can interact with both cognitive load and learning outcomes.
Socially shared regulation of learning(SSRL) is closely related to the quality of learning. However, it is difficult for SSRL in CSCL to occur autonomously, as it often requires the support of certain strategies and tools. This study aimed to introduce Artificial intelligence(AI)-Enhanced Group Awareness tools(GATs) and Adaptive Prompts (APs) as adaptive SSRL strategies into CSCL.In this study, a 2 x 2 quasi-experiment was conducted with 64 science pre-service teachers, equally divided into three experimental groups and one control group. The results found that: (1)The combination of the two strategies was able to comprehensively affect the SSRL process. (2) In terms of improving the level of SSRL and instructional design skills, the combination of the two strategies was more effective than AI-Enhanced GATs on their own, which was more effective than AI-enhanced APs on their own.
The combination of AI and plagiarism is an emerging issue following the coining of the term AI-giarism. However, there has been little research that investigated the factors that lead students to engage in AI-giarism. In response to this gap, the present study adopts the fraud triangle framework to examine students' intentions toward AI-giarism and identify the underlying factors contributing to it. Data were collected from 312 students enrolled in 25 universities and analyzed using structural equation modelling. The results indicate that AI capacity, Justification of plagiarism, unawareness of AI deception, and academic pressure increase AI-giarism behaviour among students. In contrast to previous research, the study found no significant relationship between AI-giarism and either lax enforcement or a lack of understanding of AI. By offering empirical insights into the antecedents of AI-giarism, the present study advances the current body of literature, which has been more conceptual or student perception-centric.
Peer interactions in higher education settings frequently occur in informal, private student chat groups. However, little is known about how co-regulatory learning interactions in these groups influence students' self-regulated learning, a process that is positively associated with academic success. The aim of this qualitative study was to explore the nature of the relationship between co- and self-regulated learning within the context of student-initiated online chat groups. Data were collected through semi-structured interviews and analysed utilising a reflexive thematic analysis framework. Five overarching themes were identified in the data: (a) flexible peer connection, (b) knowledge sharing, (c) monitoring progress, (d) support and reassurance, and (e) motivation. The findings suggest that students regularly engage in co-regulatory learning interactions with their peers, motivated by convenient, flexible access to timely support. Group chats were perceived as valuable environments where social interactions helped students achieve their individual learning goals. The novel results of this study help refine and advance theoretical understandings of co-regulated learning and its relationship with self-regulated learning. They offer initial insight into the mechanisms by which co-regulated learning may facilitate the emergence of self-regulated learning and highlight the important role social context plays in shaping regulatory learning interactions. The findings suggest productive areas for further research to corroborate, expand and quantify these exploratory insights.
In the evolving landscape of higher education, metaverse is emerging as a transformative force. This study explores the experiences of 388 university students across two “Soft skills” courses, one using a traditional blended learning model and the other integrating the metaverse platform GatherTown. Results indicate that GatherTown was associated with better Time Management and Task Strategies and provided contextual features that interacted with learners' self-regulated learning to support engagement, rather than directly improving test performance. Qualitative feedback emphasized peer collaboration as a key benefit of the metaverse environment and outlined students' expectations for intuitive, user-friendly AI virtual agents capable of providing personalized feedback and adaptive support. Together, these findings highlight the potential of metaverse platforms in higher education and offer initial insights into designing effective AI agents that effectively enhance self-regulated learning.
The question of how to use artificial intelligence generated content (AIGC) properly to enhance learning among college students is a key concern for contemporary educators. Although previous studies have discussed the influence of AIGC on college teaching and student learning and its functions in this context, there remains a lack of discussions regarding ways of guiding students' use of AIGC and studies on the specific topic of helping college freshmen use AIGC properly. Based on the substitution, augmentation, modification and redefinition (SAMR) model, this study develops a progressively active teaching framework that integrates AIGC into learning. This framework is used to design learning activities for general education courses targeting freshmen. This exploratory study was conducted in the context of a 16-week course. During the teaching process, AIGC interaction log data and AIGC experience records were collected from students, following which data processing was conducted using the discourse analysis, quantitative statistical analysis, and epistemic network analysis (ENA) methods to obtain the ultimate results of this study: (1) A combination of active teaching with the SAMR model can improve the quality of interactions between students and AIGC; (2) teaching strategies rooted in active learning can enhance students' ability to use AIGC; and (3) improvements in students' technical skills strengthen the quality of their interactions with AIGC. This study makes novel contributions to the literature on active learning strategies for teachers and curriculum designers, and it offers practical guidance for educational practitioners and college students regarding the integration of AI technology into both teaching and learning.
