Generative AI (GenAI) is rapidly transforming higher education. While textual and visual AI tools are increasingly used as learning assistants and creative supports, limited research has examined the competencies students develop, the interaction patterns that emerge, and the emotional dynamics shaping these experiences. This qualitative study investigates how university students interact with GenAI in academic contexts and the emerging forms of learner–AI partnerships. Semi-structured interviews combined with guided observation demonstrations were conducted with 19 undergraduate and graduate students from multiple disciplines. Using an inductive thematic analysis, the study identified three key dimensions of learner–GenAI interaction. First, students reported developing a range of AI-related competencies, including prompting practices, contextual and decomposition strategies, mental verbalization, and critical evaluation of AI outputs. Second, three partnership types emerged from the data: AI-led, Learner-led, and Collaborative interactions, reflecting different levels of learner control and cognitive engagement when working with AI systems. Third, students frequently expressed anthropomorphic behaviors, such as courtesy, companionship, and role-play, indicating that emotional and relational dynamics shape learner–AI interaction. Building on these findings and grounded in the present qualitative dataset, the study proposes a conceptual Partnership Model of Learner–GenAI Collaboration that integrates partnership types, skill development, and relational engagement. The findings contribute to emerging discussions on AI literacy by highlighting competencies and interaction patterns involved in human–AI collaboration in educational settings.
Existing Community of Inquiry research has not fully explored how social and cognitive processes jointly shape conceptual development in asynchronous online discussions. To address this gap, this study analyzed the semester-long discussion of 54 graduate students enrolled in a fully online educational foundations course. We employed a layered analytical approach integrating structural topic modeling to identify conceptual themes, qualitative content analysis to determine cognitive phases, social network analysis to measure students' social centrality, and epistemic network analysis to model structural changes in discourse over a semester. This multi-method approach allowed for an integrated examination of how conceptual, cognitive, and social dimensions co-evolve in graduate-level discussions. We identified two primary conceptual domains (Classroom Practice and Sociocultural Factors) and a nonlinear epistemic trajectory characterized by an “Expansion-Return” pattern: initial grounding in classroom practice, an outward expansion toward systemic sociocultural analysis mid-semester, and a structural return to immediate practice concerns at the end of the course. The results revealed that while initial phases exhibited structural inertia, major conceptual restructuring emerged primarily during the Resolution phase. Furthermore, students with higher social centrality demonstrated more integrated and holistic networks, connecting systemic and practice-oriented concepts more extensively than their lower-centrality peers. These findings provide important insights into how conceptual change unfolds in online environments and suggest that instructional design may benefit from targeting the transition to the Resolution phase to support effective epistemic restructuring.
Learning analytics dashboards (LADs) visualize learner data in Internet-enabled higher education, but meta-analytic evidence on the effect of the dashboard itself is scarce. Following PRISMA 2020, five databases were searched for experimental and quasi-experimental studies (January 2013 to March 2026). After verification against the primary reports, 22 studies met the inclusion criteria (N = 33,409), and 21 were pooled. Dashboards embedded in adaptive tutoring or intensive automated feedback were excluded. Student-facing and teacher-facing dashboards were synthesized separately as distinct causal pathways (g = 0.287, k = 16, and g = 0.260, k = 5); the teacher-facing estimate is fragile, resting on non-randomized studies and one very large study. The estimate across all 21 studies is reported only as a descriptive baseline of the literature, not an expected return (Hedges' g = 0.276, 95% CI [0.162, 0.390]); it was g = 0.295 in higher education samples and g = 0.219 in randomized trials. Reinstating two confounded systems inflated the baseline to g = 0.461. Heterogeneity was high (I2 = 79.2%), and the 95% prediction interval [−0.161, 0.712] spanned zero, so an implementation may yield no benefit. Effects were significant for academic performance (g = 0.237) and completion (g = 0.122), but not for log-based engagement, self-reported self-regulation, or motivation. No design, audience, era, or modality moderator was significant, and trim-and-fill suggested a bias-adjusted mean near g = 0.17. Institutions should treat LADs as implementation-contingent: pilot with usage telemetry, use normative comparison with care, and evaluate before scaling.
Although prior research has demonstrated the benefits of prequestions for learning, most studies have focused on overall effects or individual design features, leaving the interaction between question depth and placement underexplored. The present study employed a 2 (depth: deep vs. shallow) & times; 2 (placement: before learning vs. interspersed) between-subjects design. One hundred and twenty college students were randomly assigned to four conditions and learned about the fundamental characteristics of perception through a video lecture. Retention and near transfer tests were used to assess learning outcomes, and cognitive load was also measured. Results indicated that interspersed prequestions significantly enhanced retention performance compared to prequestions presented before learning, whereas deep prequestions significantly enhanced near transfer performance compared to shallow prequestions. A significant interaction between depth and placement was observed for near transfer. The advantage of deep over shallow questions emerged only when the questions were interspersed, and the advantage of interspersed over before-learning placement emerged only for deep questions. No significant differences in cognitive load were observed across conditions. These findings highlight the importance of considering both design features when implementing prequestions and offer preliminary guidance for the design of prequestions in video-based instruction.
Learning analytics has been widely applied to examine student engagement and academic performance in online learning environments, yet most existing frameworks treat engagement-performance relationships as context-independent. This study challenges that assumption by adopting a context-sensitive learning analytics perspective, empirically demonstrating that engagement patterns and their academic consequences are systematically shaped by course-level contextual configurations. An explorative, correlational, and comparative design was employed, using LMS log data from one of Southeast Asia's largest open and distance education institutions. Data from 52 courses involving 5044 students were analyzed using descriptive statistics, Pearson correlation analysis, and a General Linear Model-based one-way Multivariate Analysis of Variance (MANOVA). An Intraclass Correlation Coefficient (ICC = 0.343, 95% CI [0.325, 0.361]) confirmed that 34.3% of variance in academic performance is attributable to between-class factors, providing substantial empirical support for a context-sensitive analytical approach. The findings reveal three novel patterns: (1) the direction of discipline-based engagement differences reverses as a function of course size; (2) tutor discussion activity and student discussion scores are systematically decoupled in large courses; and (3) course category exerts a significant contextual effect on final performance (Pillai's Trace F(15,14,334) = 5.208, p < .001; Final Score: F(5,4778) = 4.121, p = .001; post-hoc: social science courses in large and middle-size settings outperform STEM-oriented courses across all three simultaneous performance outcomes). These findings demonstrate that decontextualized analytics frameworks may be inadequate for large-scale open and distance education settings, with direct implications for instructional design and institutional analytics policy.
