Generative artificial intelligence (GenAI) has become a routine source of feedback, explanation, drafting support and task guidance in higher education. This study examined how individual students' self-reported GenAI reliance and perceived human agency co-varied over time in an AI-supported collaborative writing course. Three-wave data were collected from 342 undergraduates working in 85 fixed groups. A traditional cross-lagged panel model (CLPM) and a random intercept cross-lagged panel model (RI-CLPM) were estimated to distinguish observed-level associations from within-person temporal relations after stable between-person differences were separated. The CLPM indicated reciprocal observed-level associations. In the RI-CLPM, higher-than-usual reliance was associated with lower subsequent agency, whereas agency-to-reliance paths were weaker and not statistically supported. Equality-constraint tests were consistent with directional asymmetry, while 95% confidence intervals and Monte Carlo sensitivity analyses showed that small reverse effects remain possible. Backend logs showed that self-reported reliance corresponded to AI-use intensity, although log indicators could not distinguish strategic consultation from cognitive offloading or deference. The findings highlight the importance of designing AI-supported collaboration so that students' responsibility for interpretation, judgement and authorship remains visible.Practitioner notes What is already known about this topic GenAI systems can provide timely feedback, explanation and drafting support in collaborative learning environments. AI support may help students generate ideas and revise work, but reliance on system-generated suggestions can reduce opportunities for judgement, negotiation and ownership. Traditional cross-lagged models can conflate stable between-person differences with within-person temporal associations. What this paper adds The study examines individual students' self-reported GenAI reliance and perceived human agency across three waves in an AI-supported collaborative writing course. After stable between-person differences were separated, RI-CLPM estimates provided stronger evidence for paths from higher-than-usual reliance to lower subsequent agency than for the reverse direction. Equality-constraint tests were consistent with directional pattern asymmetry, while confidence intervals and Monte Carlo sensitivity analyses indicated that small agency-to-reliance effects remain possible. Backend logs showed that self-reported reliance corresponded to AI-use intensity, but log indicators could not distinguish strategic consultation from cognitive offloading or deference. Implications for practice and/or policy AI-supported collaborative tasks should make students' responsibility for interpretation, judgement and authorship visible throughout the learning process. Instructors can ask groups to form an initial position before invoking AI and to explain why AI suggestions are accepted, revised or rejected. GenAI tools may better support agency by offering critics, asking questions and prompting reflection while leaving decisions about task direction to students. Learning analytics dashboards should not equate high AI-use intensity with over-reliance unless log data are interpreted alongside discourse, reflection, revision or decision-making evidence.
Young children increasingly engage with digital media tools for purposes of education, entertainment and communication. This study examines children's digital media activities within the context of their larger home learning environment (HLE). Using data from two different samples (N = 289), we looked at digital media use and found that caregivers reported that children most frequently engaged in watching content and playing literacy games compared to STEM-related activities. This more comprehensive view of the HLE aligned with three distinct factors: analog activities, digital learning and watching content. Analog activities were significantly and positively associated with children's language comprehension and vocabulary skills, whereas digital learning activities were significantly and negatively associated with their vocabulary skills. Watching content showed no significant associations with children's learning outcomes. In our data, digital learning activities, as currently used, did not confer the same benefits as analog activities; guidance on selecting high-quality tools and on supporting co-use may be required to realize digital benefits.Practitioner notes What is already known about this topic In this digital age, young children engage with digital media tools (smartphones, tablets, e-readers and apps) more often than in the past. Digital media is often used as part of children's home learning environment (HLE). There are concerns that too much digital media is negatively related to children's outcomes. What this paper adds Young children who are from low-income households are more likely to engage in watching content and playing literacy games in comparison to STEM-related activities and games. Data uncovered three factors associated with the HLE: analog activities, watching digital content and using digital learning activities. Analog activities were positively associated with children's language skills, but digital learning activities were negatively associated with their vocabulary. Watching content was not associated with any child skills. Implications for practice and/or policy Educators and families may benefit from guidance on prioritizing analog learning activities while also carefully incorporating digital activities. Future research should investigate the quality and the characteristics of the digital activities children engage in to better understand their role in supporting children's skills.
Abstract Generative AI (GenAI) is increasingly integrated into collaborative inquiry to provide metacognitive prompting that scaffolds socially shared regulation of learning (SSRL). Understanding how such GenAI‐based metacognitive prompting is taken up through student–GenAI interactions is therefore critical. However, although learner–GenAI interaction processes have been widely examined, little is known about how learners' preconceptions shape their engagement with GenAI‐based metacognitive prompting. This study examined how groups with different or similar preconceptions about energy forms and conservation differed in their engagement with GenAI‐based metacognitive prompting during collaborative inquiry. Participants comprised 14 small groups, including eight groups with different preconceptions (D groups) and six groups with similar preconceptions (S groups). The distribution of learner‐generated inquiries directed to the written metacognitive AI (MAI) was largely comparable across groups. Lag sequential analysis revealed contrasting patterns in how groups evaluated the metacognitive prompts: D groups showed sustained positive or neutral evaluation of the prompts, whereas S groups showed sustained negative evaluation. Process mining further showed that AI query negotiation served as a hub in D groups' collaboration, connecting their conceptual discussions and task coordination. In S groups, by contrast, it formed part of a simpler, linear sequence of querying the AI and processing its responses. This study advances understanding of how preconception diversity shapes the processes and dynamics of human–AI collaboration. The findings show that AI‐mediated regulation varies depending on group‐level knowledge configurations. Practitioner notes What is already known about this topic Effective collaborative inquiry depends on learners' prior knowledge resources and their regulatory capacity. Preconceptions, as a core component of prior knowledge, shape how learners interpret problems and engage in conceptual discussion. GenAI‐based metacognitive prompting is designed to externalise metacognitive awareness and scaffold learners' regulatory processes. Regulation unfolds through temporally ordered interaction sequences that reveal underlying regulatory dynamics. What this paper adds This study compares how different preconception groups and similar preconception groups engage with GenAI‐based metacognitive prompting during collaborative scientific inquiry. GenAI‐based metacognitive prompting responded to similar types of learner inquiries across groups. GenAI query interaction in different preconception groups appeared more integrated with conceptual and task discussion, whereas in similar preconception groups it followed a more linear, output‐focused path. Implications for practice and/or policy GenAI‐based metacognitive prompting may benefit from incorporating a preliminary classification layer that distinguishes between metacognitive prompts and procedural or informational requests. Teachers may consider preconception diversity when forming groups, as different epistemic configurations appear to give rise to distinct patterns of GenAI‐learner collaboration. Policymakers and designers should recognise that GenAI functions as a co‐regulator within collaborative settings, and its impact depends on how it is interactionally integrated into group interactional trajectories.
Abstract Socially shared regulation of learning (SSRL) is essential for successful collaboration, yet effectively scaffolding it remains challenging because regulatory needs are dynamic and context‐dependent. While generative AI (GenAI) offers potential for adaptive support, current applications in collaborative learning largely position it as a content‐oriented tutor for knowledge explanation or idea generation, leaving its role as a process‐oriented team moderator for scaffolding SSRL processes underexplored. To address this gap, we conducted a 16‐week quasi‐experiment with 64 undergraduates in China, comparing adaptive SSRL scaffolding delivered by a GenAI agent in the adaptive‐SSRL condition with static scaffolding via structured digital worksheets in the static‐SSRL condition. Survey data and group discussion transcripts were analysed using statistical methods and epistemic network analysis. Results indicated several differences in students' multidimensional engagement between the two conditions: (1) Behaviourally, the adaptive‐SSRL group showed significantly higher participation equality, although participation intensity did not differ significantly between groups. (2) Cognitively, the adaptive‐SSRL group demonstrated deeper cognitive engagement, reflected in more advanced cognitive presence phases and more integrative, solution‐oriented discourse patterns. (3) Emotionally, the adaptive‐SSRL group reported significantly higher overall emotional engagement, particularly feeling more relaxed during collaboration and more connected with peers. These findings extend human–AI collaboration research by showing that GenAI can function as an SSRL‐informed team moderator rather than a content‐oriented tutor. They also offer design implications for GenAI‐enabled adaptive SSRL scaffolding and human–AI role division in CSCL. Practitioner notes What is already known about this topic Socially shared regulation of learning (SSRL) is critical for successful collaboration and is closely associated with the quality of student engagement. While static scaffolds (eg, worksheets) provide a basic regulatory structure, they fail to adapt to the dynamic and emergent needs of collaborating teams. Generative AI (GenAI) is increasingly used for personalised content support, yet its potential as a process‐oriented team moderator remains underexplored. What this paper adds This study introduced SSRLAgent, a GenAI agent grounded in SSRL theory and established principles of collective metacognitive support, to provide adaptive and interactive regulatory scaffolding. In a 16‐week quasi‐experiment, we compared GenAI‐enabled adaptive SSRL scaffolding (SSRLAgent) with static scaffolding (structured digital worksheets) in CSCL. Adaptive‐SSRL scaffolding was associated with higher‐quality engagement: greater participation equality, deeper cognitive engagement and higher overall emotional engagement. Implications for practice and/or policy The study underscores the necessity of shifting SSRL scaffolding from static, ‘one‐size‐fits‐all’ support to adaptive GenAI interventions to improve collaborative engagement. The findings provide design guidance for developing SSRL agents that are embedded in the shared interaction space and responsive to groups' moment‐to‐moment regulatory needs. The study suggests a productive human–AI division of labour: GenAI agents can support process coordination and regulatory prompting, while teachers and peers remain essential for relationally complex forms of support that collaborative learning requires.
Abstract While generative AI (GenAI) tools are emerging as educational resources, general‐purpose applications like ChatGPT are not explicitly designed to support the metacognitive and socio‐emotional dimensions of learning, often failing to address students' emotional states or promote independent thinking. In response, this study introduces a CVT‐SRL GenAI Support Framework integrating control‐value theory (CVT) and self‐regulated learning (SRL) and tests a learner‐centred GenAI chatbot design employing prompt‐based SRL scaffolding and emotional support. In a preregistered experiment in an authentic educational setting, 144 high school students were randomly assigned to one of three conditions: (1) ChatGPT, (2) SRLbot (GenAI with SRL scaffolding), (3) emotional SRLbot (GenAI with SRL and emotional support). Outcomes included conceptual knowledge and transfer, trace‐based and self‐reported engagement in SRL phases (forethought, performance, self‐reflection), enjoyment, boredom, hopelessness, self‐efficacy and trust. Data were collected at pretest, posttest and 4‐week follow‐up. Across conditions, students showed improvements in conceptual knowledge and transfer from pretest to posttest, with no significant condition differences in post‐to‐follow‐up change. Both SRLbot and Emotional SRLbot were associated with higher self‐reported enjoyment and lower boredom than ChatGPT and elicited trace‐based SRL behaviours. Moreover, Emotional SRLbot prompted greater engagement in performance‐phase and self‐reflection processes than the other conditions. These findings provide preliminary evidence that GenAI agents integrating SRL and emotional scaffolding can support more positive emotional states and task‐level SRL behaviours than general‐purpose GenAI, highlighting control‐value appraisals as a promising target for emotionally supportive SRL design. Future research should explore longitudinal effects on SRL skill development and independent regulation. Practitioner notes What is already known about this topic Generative AI (GenAI) tools such as ChatGPT have shown promise in enhancing learning outcomes, but they also present challenges in supporting socio‐emotional needs, fostering metacognition and promoting higher‐order thinking. Applying self‐regulated learning (SRL) frameworks in chatbot‐assisted learning has been found to foster learner autonomy, long‐term engagement and improved learning outcomes. Achievement emotions experienced during learning play a critical role in supporting self‐regulated learning, motivation and academic achievement. Interventions that target learners' perceptions of control and value through reappraisal can effectively foster positive emotions. What this paper adds An investigation of the impact of three GenAI conversational agents (standard ChatGPT, an SRL agent and an emotionally supportive SRL agent) on cognitive and socio‐emotional outcomes as well as task‐level self‐regulatory behaviours among high school students in an authentic classroom setting. Preliminary evidence that integrating prompts which target self‐regulated learning phases and learners' emotional states, as well as address perceptions of control and value can increase students' enjoyment of concrete learning situations and reduce boredom more effectively than standard ChatGPT. Insights into how theoretically grounded GenAI agents that incorporate SRL and emotional scaffolding may support students' emotional engagement and task‐level self‐regulatory behaviours during a short classroom learning activity. Implications for practice and/or policy Prompt engineering can be used to simulate emotional support and personalize guidance based on SRL and emotion regulation principles, providing design considerations for future learner‐centred GenAI systems. Developers and educators aiming to design pedagogically meaningful GenAI tools should focus on scaffolding SRL both cognitively and emotionally, to support the affective and self‐regulatory dimensions of learning. We make the technical specifications and system prompts publicly available.
