Abstract Despite the increasing application of AI chatbots, few studies have examined the effectiveness of AI chatbot‐supported cooperative‐flipped classrooms in university teaching. Based on mastery learning theory, this study employed a quasi‑experimental design to examine how an AI chatbot‐supported cooperative‑flipped classroom influences students' collaboration, self‐regulated learning and academic performance, and whether students with different prior knowledge levels demonstrate distinct patterns under AI‐supported flipped learning. This study involved 154 junior students from a normal university, including an experimental group ( n = 87) and a control group ( n = 67), who were taught by the same instructor over an 11‐week period. Quantitative data from pretest and posttest scores and self‑report scales were analysed using t ‐tests and ANCOVA, whereas qualitative interview data from 10 experimental group students were analysed using Epistemic Network Analysis to compare collaboration and self‑regulated learning patterns across prior knowledge groups. The results showed that, compared to the control group, the experimental group demonstrated significantly higher posttest scores in collaboration, self‐regulated learning and academic performance. Within the experimental group, students with higher levels of AI chatbot interaction also significantly outperformed those with lower interaction levels in academic performance. Further, Epistemic Network Analysis results revealed that in AI chatbot‑supported flipped learning, students with lower prior knowledge exhibited denser collaboration networks, requiring more cognitive input and more frequent dimension switching to coordinate collaborative processes. In contrast, students with higher prior knowledge demonstrated stronger connections between the seeking, engaging and reflecting dimensions in their self‐regulated learning networks, reflecting more integrated self‐regulatory behaviour. This study provides empirical support for mastery learning theory and demonstrates the effectiveness of AI chatbot‐supported cooperative‐flipped classrooms, offering implications for differentiated teaching based on students' prior knowledge levels. Practitioner notes What is already known about this topic Flipped classrooms are associated with improvements in students' self‐regulated learning, motivation and academic performance compared with traditional instruction. Group cooperation during pre‐class activities can enhance peer interaction and collaboration, but is often constrained by uneven participation and limited instructional support. AI chatbots have been applied as learning assistants in higher education to provide feedback and learning resources, yet their role in structured pre‐class group cooperation remains insufficiently examined. What this paper adds AI chatbot‐supported cooperative‐flipped classrooms significantly enhance students' collaboration, self‐regulated learning and academic performance compared with traditional group‐based flipped learning. Epistemic network analysis shows that students with different levels of prior knowledge exhibit distinct collaboration and self‐regulation patterns when supported by an AI chatbot. The findings demonstrate that AI chatbots can function as mastery‐oriented scaffolds by guiding task division, clarifying learning goals and supporting feedback during pre‐class group discussions, thereby strengthening students' readiness for subsequent in‐class learning. Implications for practice and/or policy Practitioners are encouraged to integrate AI chatbots into pre‐class group discussions to provide structured guidance, timely feedback and support for collaborative knowledge construction. Instructional support should be differentiated by prior knowledge: learners with lower prior knowledge benefit from clearer goals and stronger cognitive scaffolding, whereas higher‐achieving learners benefit from tasks emphasizing integration, reflection and assessment. For large‐scale or resource‐constrained teaching contexts, AI chatbot‐supported cooperative‐flipped classrooms offer a scalable approach to implementing mastery learning and differentiated instruction in technology‐enhanced pedagogy.
Abstract Emotion regulation has been recognized as a key factor affecting students' academic success. However, in conventional school settings, it is challenging to foster students' emotion regulation owing to the involvement of complex factors that might influence students' conscious and unconscious behaviours. To address this issue, a self‐determination theory (SDT)‐based emotional agent framework was proposed, and an emotional agent (EmoAgent), capable of proactively detecting students' emotional states and providing emotional regulation strategies to individual students in addition to the provision of conventional learning supports, was implemented. To assess the effectiveness of the proposed approach, an 8‐week quasi‐experiment was designed with 173 sixth graders from four classes taking the Information Technology course in a primary school. Two classes (87 students) comprised the experimental group (EG) which used the SDT‐based emotional agent‐mediated (SDT‐EA) approach, while the other two classes (86 students) comprised the control group (CG) which learned with the SDT‐based conventional agent (STD‐CA), which is capable of providing learning supports but without considering emotional factors. The experimental results showed that (1) the SDT‐EA approach significantly outperformed the SDT‐CA approach in academic achievement; (2) the SDT‐EA approach engaged students in positive emotional experiences, whereas the SDT‐CA approach exposed students to more negative emotional experiences; (3) negative emotion was significantly negatively associated with academic achievement, moderated by the SDT‐EA approach. Practitioner notes What is already known about this topic? Within self‐determination theory, if the three psychological needs of autonomy, competence and relatedness are met, students’ intrinsic motivation is activated and in turn their academic emotion is regulated, which shapes their well‐being, engagement and academic success. GenAI–human collaboration can facilitate students’ self‐determined and personalized learning. Most pedagogical agents tend to assist with personalized cognitive support rather than with emotion regulation or emotional support. What does this paper add? Merging affective computing technology with GenAI, this study constructed an emotional agent with emotion sensing, recognition, feedback and regulation, providing emotional and cognitive supports. Using process mining techniques, this study found that the emotional agent facilitated students’ emotion regulation from negative to positive. The emotional agent moderated the effect of negative emotions on academic success. The emotional agent can significantly promote academic achievement. Implications for practice and/or policy The study findings imply that the SDT‐based emotional agent‐mediated approach shows promise for strengthening students’ self‐regulated learning. The study offers a viable model for integrating socio‐emotional learning with subject‐matter instruction. The SDT‐based emotional agent can be embedded in regular instruction to instantiate a self‐determined learning paradigm.
Abstract This meta‐analysis systematically examines 35 empirical studies (2013–2025) investigating artificial intelligence applications within Zimmerman's cyclical model of self‐regulated learning (SRL). Three principal discoveries emerge: (1) Technological progression has evolved through three co‐existing paradigms: rule‐based architectures, data‐driven adaptive systems and generative AI ecosystems, that demonstrate increasingly sophisticated capabilities for human‐AI collaboration. (2) While AI‐supported SRL interventions yield a moderate overall effect size ( g = 0.507), their impact is uneven; AI is significantly more effective during the task performance phase ( g = 0.574) than in the preparatory forethought phase ( g = 0.401). Notably, generative AI shows markedly superior efficacy across all phases (e.g. g = 0.709 for forethought, g = 0.938 for performance), though high heterogeneity suggests these effects are heavily contingent on specific instructional designs. (3) Moderator analysis identifies optimal contexts in secondary education, natural science disciplines, fully online settings and interventions of medium duration (2–10 weeks), while also revealing that effects are substantially larger when measured by behavioural traces compared to self‐reports. Critically, these findings highlight a persistent performance‐competence divide, suggesting that AI's capacity to scaffold immediate task performance may outpace its current ability to cultivate durable, transferable self‐regulatory competence. The study discusses the implications of this divide and proposes a research agenda focused on designing AI systems that foster genuine learner autonomy. Practitioner notes What is already known about this topic AI technologies have shown promise in supporting self‐regulated learning (SRL) through personalized feedback and metacognitive scaffolding. Previous studies report inconsistent outcomes of AI interventions across SRL phases (forethought, performance, self‐reflection), with limited exploration of phase‐specific impacts Existing meta‐analyses often treat SRL as a unified construct rather than examining how different AI types support distinct self‐regulatory processes. What this paper adds Demonstrates moderate overall effectiveness of AI‐supported SRL ( g = 0.507) with differential impacts across phases: strongest during task performance ( g = 0.574), weaker in forethought ( g = 0.401) and self‐reflection ( g = 0.464) Maps three coexisting AI paradigms—rule‐based, data‐driven and generative AI—revealing that generative AI achieves superior outcomes across all SRL phases, though effectiveness depends on alignment between AI affordances and specific self‐regulatory processes. Identifies optimal implementation contexts through moderator analysis: secondary education settings, natural science disciplines, fully online environments and medium‐duration interventions (2–10 weeks) yield stronger effects. Articulates a critical performance‐competence divide, substantiated by quantifying how AI's impact on observable behaviours ( g = 0.751) is more than double its effect on learners' self‐reported perceptions ( g = 0.369). Implications for practice and/or policy Integrate AI paradigms strategically rather than viewing them as replacements: combine rule‐based systems' stable scaffolding with generative AI's adaptive dialogue to support the full SRL cycle. Design interventions of medium duration (2–10 weeks) to optimize skill acquisition while maintaining engagement, avoiding both novelty effects and scaffold dependency. Adapt AI implementation to disciplinary demands: structured procedural support works well in natural sciences, while social sciences require enhanced support for open‐ended inquiry and critical analysis. Develop assessment frameworks that measure delayed, unsupported performance alongside immediate gains to ensure AI fosters genuine self‐regulatory competence rather than temporary performance enhancement.
Abstract With the increasing application of GenAI in education, researchers and practitioners are paying more and more attention to its effectiveness and impact in teaching and learning. This can be evidenced by the increasing number of literature reviews on this topic published in recent years. However, these literature reviews seldom analysed the effectiveness of GenAI‐powered educational applications from the human–AI interaction perspective, which is a widely recognized critical factor influencing the effectiveness of such GenAI‐powered educational applications. In response, this study systematically reviews 56 empirical studies on the application of GenAI in education. To explicitly address this gap from the human–AI interaction perspective, we analysed the reviewed studies using the AIED‐HCD framework, which conceptualizes three human–AI interaction modes along the dimensions of human control and AI automation, and examined educational contextual factors and educational tasks supported by GenAI to assess how interaction modes vary across teaching and learning contexts. In addition, we conducted a sensitivity analysis to evaluate the robustness of findings across these modes. We demonstrated that, although current educational practices remain cautious towards interaction modes with a high level of AI automation, the mode characterized by both high human control and high AI automation has begun to emerge as a trend, demonstrating promising potential for integrating the respective strengths of humans and AI. Furthermore, sensitivity analysis reveals that many studies lack sufficient detail in their statistical reporting, and the reported effect sizes often fall below the thresholds required for acceptable statistical power. Based on these findings, we recommend: (i) beyond exploring how to improve the practical use of AI automation under human supervision and control, educational researchers and practitioners should carefully choose and implement suitable human–AI interaction settings according to the specific context of use, with higher levels of AI automation applied only when supported by appropriate task design and pedagogical guidance; (ii) researchers should improve methodological transparency by estimating appropriate sample sizes and testing assumptions to ensure the reliability of empirical findings. Practitioner notes What is already known about this topic Generative artificial intelligence can efficiently analyse vast amounts of textual information and perform complex natural language processing and generation tasks, demonstrating powerful language intelligence capabilities. Generative artificial intelligence is increasingly being integrated into various educational systems, and with its exceptional capabilities, its potential to support educational applications is gradually being explored and put into practice. What this paper adds Based on the AIED‐HCD conceptual framework, this study systematically classifies current GenAI‐based empirical research by examining two key dimensions, namely the extent to which human control and automation through GenAI are enabled, and offers a holistic perspective on the current state of research. A sensitivity analysis was conducted to examine whether the empirical evidence reported in current GenAI‐based studies demonstrates sufficient statistical power to support future research and practical implementation. Implications for practice and/or policy Stay continuously informed about the latest advancements in AI technologies and carefully verify their suitability for different interaction modes to enhance educational user experiences and deliver systematically measurable improvements and efficiencies to the educational landscape. Current GenAI‐based empirical studies should include more detailed statistical reporting, such as whether assumption tests were conducted and the specific results of effect sizes, to support empirical evidence and enhance transparency.
Abstract Digital game‐based learning often leverages social dynamics to enhance engagement, yet the interplay between collaborative structures and competitive pressures remains complex. This study investigated different collaboration modes and how they interact with competition to influence computational thinking (CT) learning, group metacognition and in‐game behaviours. Adopting a quasi‐experimental approach with a 3 × 2 factorial design, 148 seventh‐grade students were recruited via purposive sampling from intact classes. Participants played the game in pairs, with or without competition, in one of three collaboration modes: paired gaming on one computer, collaboration on respective computers with turn‐taking or cooperation on respective computers with role assignment. Data were analysed using quantitative methods to assess learning outcomes and behavioural sequence analysis to identify interaction patterns. Results revealed significant main effects of collaboration mode and an interaction effect between collaboration and competition. The cooperative group with competition achieved the highest CT outcomes, and the paired gaming group with competition reported the highest level of group metacognition. In contrast, the collaborative group with competition had the lowest level of learning outcomes and group metacognition. Behavioural sequence analyses identified unique patterns that corroborate the quantitative findings. Notably, competition undermined the collaborative turn‐taking mode while enhancing cooperative and paired gaming modes, highlighting the need to carefully match collaboration structures with competitive elements in designing game‐based learning. Practitioner notes What is already known about this topic Collaboration and competition are widely used in digital game‐based learning (DGBL) to enhance engagement, motivation and problem‐solving. Prior studies have often examined collaboration and competition separately, leaving the interactive effects between these social dynamics on computational thinking (CT) and group metacognition underexplored. What this paper adds Different collaboration modes significantly shaped CT learning and group metacognition. Competition alone showed no benefits for CT learning or group metacognition. Cooperative + competition fostered clear roles, strategic play and higher CT. Competition weakened collaboration learning's CT performance and led to fragmented behaviours. Implications for practice and policy Carefully match collaboration structures with competition features when designing game‐based or group learning activities. Use competition selectively—beneficial in cooperative tasks with clear role assignments, but potentially detrimental in open collaboration requiring trust and mutual regulation. Integrate competitive and collaborative elements intentionally rather than additively.
Abstract The rapid growth of massive online learning has intensified interest in self‐regulated learning (SRL) and socially shared regulation of learning (SSRL), yet empirical insights into their interplay in online collaborative learning (OCL) remain limited. This study employed a three‐layer analytical method combining cluster analysis, content analysis and complex network analysis to investigate how SSRL patterns evolve in relation to individual SRL profiles. Data from 60 undergraduates in a 16‐week course included over 16,000 trace entries (e.g. posts, replies, peer feedback) were collected and analysed. Results revealed that (1) three profiles of SRL were identified, based on learners' time investment, study regularity and help‐seeking behaviours; (2) groups with higher SSRL behavioural interaction displayed a more diverse and balanced role composition; and (3) distinct SSRL patterns emerged across SRL profiles over time. Individuals with high SRL profiles more frequently connected idea sharing with strategy using and process monitoring, while those with low SRL profiles relied on limited strategies focused mainly on idea sharing. These findings deepen the understanding of how individual regulation shapes group dynamics in online collaboration and suggest that instructional design should consider learners' SRL profiles when scaffolding their collaborative regulation processes. Practitioner Notes What is already known about this topic Self‐regulated learning (SRL) and socially shared regulation of learning (SSRL) are both essential for effective online collaborative learning (OCL), with prior research recognising their theoretical interplay. Trace data (e.g. login frequency, timestamp and posts) offer valuable opportunities to uncover how learners engage in and coordinate regulatory behaviours throughout online collaboration. What this paper adds Integrating SRL profiling with mixed network analysis, this study shows how individual SRL profiles drive divergent SSRL development, highlighting shared coordination patterns and evolving interactive roles across phases. Groups with higher SSRL interaction exhibited more diverse and balanced role composition, with central roles (e.g. leaders, animators) facilitating stronger collaborative regulation processes. Time‐series epistemic network analysis (ENA) provides a novel analytical lens to track the evolution of SSRL patterns across collaborative phases, linking SRL profiles with group‐level regulatory dynamics. Implications for practice and/or policy Tailoring support to SRL profiles allows educators to offer targeted scaffolds, using prompts and planning tools for low SRL learners, while encouraging high SRL learners to lead coordination and monitoring in SSRL. Role‐driven group design promotes collaboration structured by learners' SRL profiles.
Abstract This Special Section examines the role of implementation research in bridging the gap between educational technology research and practice within low‐ and middle‐income countries (LMICs). Implementation research, focused on the real‐world application of policies and programmes, emphasises context and stakeholder engagement. Despite its relative novelty in EdTech, this methodological framework facilitates insights into the scalability of interventions in equitable ways. Included manuscripts provide rigorous evidence addressing the opportunities and challenges of implementation research in LMICs. By establishing current understanding and initiating a dialogue about knowledge gaps and future directions, this collection establishes a robust foundation for EdTech implementation research in LMICs and beyond to inform global educational policy and practice.
Abstract While emotional design in educational AI is often presented as a universal benefit, this study challenges that assumption, investigating when, for whom and how it impacts L2 vocabulary learning. This paper reports on the first phase of a larger project, analysing data from a quasi‐experimental study with 147 pre‐service teachers who interacted with either an emotional or a regular AI agent. Data included pre‐/post‐tests, questionnaires and complete AI chat histories. Quantitative analysis revealed no overall difference in vocabulary acquisition or other affective variables. A more nuanced pattern emerged in an exploratory subgroup analysis: contrary to prevailing assumptions, the regular agent—not the emotional one—preserved significantly better learning attitudes ( p = 0.001) and intrinsic motivation ( p = 0.038) for learners with lower baseline proficiency. A possible mechanism is suggested by a significant negative correlation between self‐reported cognitive load and vocabulary scores, found only in the emotional AI group: under high task difficulty, the emotional agent's verbose scaffolding may have added extraneous cognitive load rather than serving as a supportive buffer. We treat this correlational evidence as tentative rather than demonstrated. To explain these divergent outcomes, a qualitative analysis of interaction patterns identified four distinct learner archetypes—from the high‐agency ‘Constructive Inquirer’ to the passive ‘Attentive Pupil’. The findings indicate that the emotional AI's efficacy might be moderated by these profiles; its verbose emotional scaffolding may have added extraneous processing demands for the vulnerable ‘Attentive Pupil’ while being perceived as an inefficient frustration by the task‐oriented ‘Demanding Critic’. The study concludes that a one‐size‐fits‐all approach to emotional AI is suboptimal. Practitioner notes What is already known about this topic Emotional design in digital learning environments is widely considered beneficial for enhancing learner engagement, motivation and positive emotions. Affective AI systems are increasingly being developed for education with the goal of providing personalized and psychologically supportive instruction. Individual learner differences, such as prior knowledge, motivation and confidence, are known to be significant factors that impact the effectiveness of any educational intervention. What this paper adds This paper demonstrates that the benefits of emotional AI are not universal. Its efficacy is highly conditional, with the regular agent—not the emotional one—preserving better learning attitudes and motivation for lower‐proficiency learners under high cognitive load conditions. It tentatively provides a new explanatory framework of four learner archetypes (the Constructive Inquirer, Demanding Critic, Attentive Pupil and Disengaged Bystander) based on observable interaction patterns (agency and questioning effectiveness), which explains why different learners react divergently to the same AI design. It identifies a cognitive load threshold boundary condition: under high task difficulty, verbose emotional scaffolding may exceed learners' working' memory capacity and act as an extraneous processing burden rather than affective support, consistent with experimental evidence on the moderating role of task difficulty in emotional design. It reveals a potential design tension: emotional support is not uniformly beneficial—verbose affective scaffolding intended to help a passive learner (the ‘Attentive Pupil’) may instead add extraneous cognitive load under demanding tasks, while being perceived as an inefficient frustration by a high‐agency, task‐oriented learner (the ‘Demanding Critic’). Implications for practice and/or policy The design and implementation of educational AI must move beyond a ‘one‐size‐fits‐all’ model. Practitioners should select and advocate for tools that can adapt their emotional persona based on user needs, rather than applying a single emotional design universally. The four archetypes can serve as a preliminary diagnostic tool for educators. By observing how students interact with AI, teachers can identify different learner types and provide targeted pedagogical support (e.g., encouraging more agency in an ‘Attentive Pupil’ or helping a ‘Demanding Critic’ reframe the AI as a collaborative partner). For AI developers, the central implication is the need to build adaptive systems that can first diagnose a user's archetype and then dynamically adjust their affective strategy—offering support when needed and becoming an efficient, unobtrusive tool at other times. Educational policy should encourage the development of AI that is not just technologically advanced but also pedagogically sophisticated and emotionally adaptive, ensuring that new technologies serve a diverse student population equitably and effectively.
Virtual technology holds tremendous potential for cultivating students' perspective taking ability. It contributes to diminishing intergroup conflict and enhancing social harmony by facilitating students' understanding and adoption of an alternative perspective. This study used meta‐analysis to test the overall effect of virtual technology on students' perspective taking ability. A total of 20 experimental (quasi) studies (published between 2009 and 2024) that documented either a comparison between a group adopting virtual technologies (21 effect sizes in 15 studies) and a control group or a pre–post comparison (six effect sizes in five studies) were identified through the literature search. Utilizing a random effects model to compute effect sizes, the results revealed that virtual technology exerts a medium influence on perspective taking ability (Hedges' g = 0.505), with notable heterogeneity observed across the studies. Furthermore, the moderator analysis results indicated that virtual technology is more influential (a) on adult learners, (b) when adopted to teach declarative knowledge, (c) when offering visual clues feedback, (d) when combined with inquiry‐discovery or test assessment method, (e) when accompanied by observational learning and (f) when in an experimental period lasting 30–90 minutes. These findings provide beneficial insights for future research and practical applications aimed at adopting virtual technology to cultivate students' perspective‐taking ability. Virtual technology has been extensively incorporated into diverse educational scenarios to promote students' learning performance. Virtual technology has been considered a promising method for cultivating students' perspective taking ability, although it has not yet been substantiated through meta‐analysis. The application of virtual technology should be carefully designed to maximize its effect on students' perspective‐taking ability. Virtual technology has a medium influence on students' perspective taking ability. Identifying four key moderators that significantly influence the effectiveness of virtual technology, including learner stages, knowledge types, teaching types and experimental periods. Identifying a moderator that marginally influences the effectiveness, that is, application forms. Implement visual feedback mechanisms in virtual systems to scaffold PT ability development. Combine test assessment or inquiry‐discovery teaching with observational learning for declarative knowledge in virtual systems to enhance PT ability. Provide middle students with additional aids during virtual PT tasks.
Lecture capture is ubiquitous in higher education. Lecture capture recordings are typically accompanied by automatically generated closed captions that are sometimes corrected by humans. Students self-report that they benefit from captions, and particularly human-corrected captions. However, multimedia learning research suggests that simultaneous spoken and written information impairs learning. Experimental tests of the pedagogic benefits of captions in authentic higher education learning contexts are sparse. Consequently, we block randomised 144 undergraduate psychology students to 12 sets of lecture capture recordings with uncorrected, corrected or no captions then assessed their understanding of these materials with 12 quizzes. Our pre-registered analyses indicated that the differences between conditions were trivial (M eta 2 = 0.01; |M| delta = 0.12) and that the captioning conditions to which students were randomised did not influence their quiz performance. Follow-up surveys re-confirmed that students believe captions are beneficial for multiple reasons (eg, improved comprehension and focus; study efficiency; accessibility), and prefer corrected over uncorrected captions, even though they do not always notice the difference. As they do not impair learning, do meet accessibility requirements and are valued by students, automatically generated captions should be used routinely. However, correcting them may not reflect an optimal use of resources.
The generative AI (GenAI) has exhibited diverse functions in supporting academic writing, especially for university students who encounter both disciplinary content and academic English challenges in English medium instruction (EMI). However, existing studies have rarely addressed EMI students' behavioural intentions to use GenAI for academic writing beyond formal classroom settings. By adopting the Will, Skill, Tool (WST) model, this study explores how EMI students' intentions are shaped by emerging AI-specific variables, including perceived risks, perceived importance of policy (PIP) and AI literacy. Survey data were collected from 512 university students at an EMI university in China. The structural equation model explained 56% of the variance in EMI students' behavioural intentions. The findings indicated that will-related factors (eg, attitudes, perceived risks, PIP) and the skill factor (AI literacy) were significant antecedents of students' intentions, whereas the tool component (facilitating conditions) was not. Attitudes and perceived risks served as significant indirect pathways through which PIP and AI literacy were associated with behavioural intentions. The findings have implications for pedagogical practices and policy design in promoting EMI students' adoption of GenAI for academic writing.Practitioner notes What is already known about this topic Some Chinese university students use generative AI (GenAI) tools to address both disciplinary and linguistic challenges in academic writing outside English medium instruction (EMI) classrooms. Students' perceived risks, perceived importance of policy and AI literacy have emerged as key predictors of their adoption of GenAI, but the interplay among these variables remains underexplored. Despite the wide application in teachers' technology adoption, the Will, Skill, Tool (WST) model has rarely been adopted to explain students' intention to use GenAI in informal learning contexts. What this paper adds The WST-based model explained 56% of the variance of Chinese EMI students' behavioural intention to use GenAI for academic writing in informal settings. The will and skill components significantly predicted students' behavioural intentions, whereas the tool element did not. Both attitudes and perceived risks constituted significant indirect mechanisms linking students' perceived importance of policy and AI literacy to behavioural intentions. Implications for practice and/or policy In EMI lessons, teachers can present the benefits of GenAI for addressing both academic content and language issues in academic writing, thus building EMI students' positive attitudes towards this technology. Policymakers should develop clear guidelines on the ethical and responsible use of GenAI in academic writing to alleviate students' concerns about academic misconduct and integrity. EMI lecturers should prioritise students' development of AI literacy by demonstrating effective strategies and critical evaluation skills for students' use of GenAI in academic writing.
As students learn and practice new skills in university courses, their behaviour can change in response to competing demands and increasing content complexity. However, most metrics used to evaluate study behaviour focus on the number or sequence of activities rather than on the change of behaviour. To address this, we replicate and extend a complex dynamical systems approach to characterise recurrence in behavioural patterns and whether it changes.Using assessment logs from 1362 students in the first 5 weeks of a semester-long programming course, we examine whether changes in the patterns of formative assessment submissions can differentiate student sub-groups and predict their performance. We identify three student profiles of behavioural change. We find that higher entropy of recurrence in assessment submission patterns is associated with better performance, and that changes in this entropy signal upcoming changes in performance. We also show that higher entropy of recurrence is associated with greater timeliness of submissions. Finally, we evaluate the predictive value of early behavioural patterns and find that while student profiles of change do not outperform conventional predictive metrics, they offer complementary insights that can enable timely interpretations of student data and inform interventions. Overall, our findings extend the generalisability of behavioural metrics based on complex dynamical systems by demonstrating consistent patterns across courses, LMS types and data sources.
This study explores the impact of robot-LLM (Large Language Model) integration on collaborative creative writing, focusing on how embodiment and AI creativity influence various aspects of creative output. A total of 150 undergraduate students participated in a structured experimental design with five collaboration conditions: Human-Human (HH), Human-Computer with High-Creativity LLM (HC), Human-Robot with High-Creativity LLM (HR), Human-Robot with Low-Creativity LLM (RL) and Human-Computer with Low-Creativity LLM (CL). Creativity was assessed through expert ratings and computational analysis of originality, imagery, voice and semantic flow. The results revealed that while the Human-Robot (High-Creativity LLM) condition significantly enhanced originality, Human-Human and Human-LLM (text-based) collaborations excelled in imagery and voice. The study identified an 'embodiment paradox', where robot embodiment amplified creativity in high-creativity AI conditions, yet human collaboration remained superior in stylistic expression. Mediation analysis revealed that user engagement acted as a mediator, with embodiment compensating for low-creativity AI and amplifying the creative process with high-creativity AI. The findings have important implications for the design of collaborative AI systems, highlighting the need for a balanced integration of embodiment and AI creativity to optimize creative outcomes. This research contributes to our understanding of how human-robot-LLM collaborations can expand creative potential in writing, offering insights for future AI applications in educational and creative industries.Practitioner notes What is already known about this topic? Previous studies have explored the impact of AI in creative collaborations, with a focus on text-based models like LLMs enhancing writing quality. Embodiment in AI systems, such as humanoid robots, has been shown to affect user engagement and emotional responses, influencing creativity. Human collaboration has traditionally been seen as superior in generating stylistic elements like imagery and voice, while AI excels in originality and idea generation. What this paper adds? This research demonstrates that robot-LLM collaboration significantly boosts originality, particularly when high-creativity AI is used. The study uncovers the 'embodiment paradox', where embodied robots enhance creativity in high-creativity AI conditions but human collaboration remains superior in stylistic expression. The mediation role of user engagement is explored, showing how embodiment can enhance creative outcomes when AI creativity is low and amplify them when AI creativity is high. Implications for practice and/or policy Educators and trainers can utilize embodied AI systems in creative tasks to increase student or participant engagement and foster more original outputs. Training programmes can be structured to leverage the strengths of both human collaboration and AI, tailoring tasks based on AI's creativity levels for optimal outcomes. Policy around the integration of AI in educational and creative settings should encourage balanced AI systems that combine embodiment and creativity for enhanced collaborative work.
Artificial intelligence (AI) is increasingly shaping how young learners interact with digital technologies, yet many upper elementary students engage with AI systems passively and develop intuitive and sometimes inaccurate conceptions of how these systems work. This study examines the Foundational AI construct within a refined learning progression (LP), exploring how scaffolded instruction and dynamic assessment support conceptual shifts in students' understanding of how AI collects, learns from and uses data to make decisions. Drawing on Vygotsky's zone of proximal development and synergistic scaffolding theory, we refined the foundational AI construct of a five-level LP and designed a two-phase activity grounded in this LP to elicit and support student reasoning through structured tasks, informational scaffolds and facilitator prompts. Through mixed methods analysis of clinical interviews with 13 fourth and fifth graders (9-11 years), we identified recurring misconceptions and tracked shifts in student reasoning and movement along the Foundational AI construct of the LP. Furthermore, we examined one student's trajectory in depth to illustrate how dynamic assessment can function as a responsive instructional tool. Findings provide initial empirical insight into how scaffolded LP-aligned instruction, paired with dynamic assessment, can support young learners' movement from surface-level ideas to more structured understandings of how AI systems function. These insights contribute to the design of developmentally appropriate and contextually responsive AI learning experiences for primary education.Practitioner notes What was already known about this topic? Many young learners interact with AI technologies (eg, voice assistants, recommendation systems) but often hold surface-level or inaccurate conceptions of how AI works. AI literacy frameworks exist, but none currently provide scaffolded pathways that align with young students' developmental readiness or explicitly address their initial misconceptions What this paper adds? Provides an initial empirical examination of a refined five-level Foundational AI construct within a broader Learning Progression (LP) for upper elementary students. Demonstrates how LP-aligned scaffolded instruction, using tasks, just-in-time informational supports and decision trees, can guide students from intuitive ideas to more data-centered reasoning. Uses dynamic assessment to track and support conceptual growth, providing insight into students' readiness to reason about AI systems. Implications for practice and/or policy Scaffolded LPs that integrate structured tasks, informational prompts and dialogic facilitation can help support developmentally grounded AI instruction that is responsive to learner needs. Dynamic assessment frameworks can help researchers and educators capture students' shifts in reasoning, differentiating between ideas students can articulate independently and those requiring additional support. Designing layered, responsive scaffolds that actively elicit student reasoning and provide opportunities for reflection can support educators in guiding students' conceptual growth in AI literacy.
Anthropomorphic strategies can improve learners' performance in artificial intelligence (AI) learning environments, yet their underlying mechanisms remain unclear. Grounded in social presence theory, this study introduces the concept of identity anthropomorphism and adopts multimodal learning analytics (MMLA), combining questionnaires, electroencephalography and eye tracking to examine its effects on learning outcomes. Seventy participants completed online learning tasks under three conditions: identity-anthropomorphised AI, non-identity-anthropomorphised AI and human companionship. Results indicated that: (1) identity-anthropomorphised AI significantly improved learning outcomes compared to non-anthropomorphised AI and performed comparably to human companionship; (2) social presence and positive emotions sequentially mediated this effect, such that social presence generated a positive net impact only when it evoked positive emotions; (3) learners in the humanities and social sciences were more sensitive than science and engineering learners to the social presence evoked by identity anthropomorphism; (4) although identity anthropomorphism diverted some attention, its emotional benefits offset this disadvantage. These findings extend the mechanism of AI-induced social presence from interactive behaviours to static identity cues, support anthropomorphic design in educational AI and employ MMLA to complement learning theory. This study also highlights that the long-term deployment of such systems should account for potential risks, such as excessive emotional dependence.Practitioner notes What is already known about this topic? Online learning environments encounter challenges such as heightened learners' isolation, low engagement, and susceptibility to distraction, with social presence recognised as an important factor in mitigating these issues. Existing research on AI identity is still in its early stages, where the identity information typically assigned to AI remains relatively limited. Multimodal learning analytics (MMLA) leverages multisource data to provide more interpretable evidence for exploring complex learning processes. Effectively integrating MMLA with learning theory offers robust support for further theoretical development. What this paper adds? This study introduces the concept of identity anthropomorphism and shows that static identity cues of AI can evoke learners' sense of social presence. Identity-anthropomorphised AI companions are associated with improved online learning outcomes, with performance comparable to human companions in this study. Social presence alone does not necessarily translate into better learning outcomes; its positive effects depend on the extent to which it fosters positive emotions. Learners from the humanities and social sciences are more responsive to identity anthropomorphism than those from science and engineering. Implications for practice and/or policy Educational technology developers and instructional designers can consider incorporating identity anthropomorphism into AI learning companions. Designing human-like identity backgrounds for AI provides a relatively low-cost approach to enhancing learning outcomes in online courses. When developing AI-based learning companions, it is not sufficient to focus solely on increasing social presence. Designers should also ensure that the system supports positive emotions and helps offset attentional distractions introduced by anthropomorphic features. Policymakers and educational institutions should also consider potential unintended consequences, such as excessive emotional dependence and issues related to educational equity. These risks may be mitigated through transparent disclosure and ongoing evaluation.
EdTech Hub is at the forefront of using Implementation Research to build better evidence regarding EdTech in low- and middle-income countries (LMICs). In this contribution, David Hollow (EdTech Hub Research Director) and Verna Lalbeharie (EdTech Hub Executive Director) explore the definitions and parameters of what Implementation Research means within education, and specifically within the context of EdTech, providing practical guidance by drawing on insights from the EdTech Hub research portfolio. Two interlinked arguments are made. First, Implementation Research provides a valuable set of methodological tools for building better evidence regarding EdTech. Second, EdTech provides a useful context for understanding and improving practices about the use of Implementation Research across education. The contribution closes with reflective prompts for Implementers and Researchers when considering engaging in EdTech Implementation Research.
This study examines the critical influence of technical and structural factors-such as interface design, technical performance, payment barriers, certification access, pricing models and accessibility constraints-on learner Engagement, Retention and Inclusivity (ERI) in Massive Open Online Courses (MOOCs) and other large-scale online learning platforms. Employing a novel mixed-methods analytical method termed 2TS (Topic Modelling, Sentiment Classification, Systematic Keyword Selection and Thematic Analysis), we analysed over 226,000 user reviews from six globally recognised online learning platforms, including traditional MOOC providers and other large-scale digital learning systems: Coursera, edX, Udemy, Alison, uLesson and Khan Academy. Our findings suggest that technical instability, limited offline functionalities and unclear pricing structures have a negative impact on learner engagement and retention. Furthermore, regional payment restrictions and ambiguous subscription terms disproportionately disadvantage learners in low-resource settings, significantly undermining the inclusivity of online learning platforms. The research highlights how carefully structured platform design and robust technical infrastructure have a significant influence on digital learning outcomes. Practical implications underscore the need for platform providers to enhance platform usability, promote transparent pricing and adopt inclusive technological solutions, particularly to ensure equitable access and engagement among diverse global audiences.
