Abstract:

This study examined how different AI-supported dialogue mechanisms influence learners’ epistemic, affective, self-regulatory, and achievement development in a semester-long programming course. Students were randomly assigned to three conditions: Group 3, a structured dialogue design encouraging questioning and skepticism toward AI responses; Group 4, an agency-supportive dialogue designed to foster epistemic agency; and Group 5, a general ChatGPT-style dialogue without additional scaffolding. The intervention lasted 18 weeks, with repeated measurements collected at three phases. At each phase, students completed instruments assessing knowledge beliefs, affect, anxiety, achievement emotions, cognitive load, self-regulated learning (SRL), human–machine trust, and motivational constructs, along with course-based academic assessments.

To analyze developmental mechanisms of AI-supported learning, a layered longitudinal framework was employed, distinguishing baseline epistemic–affective dispositions, phase-sensitive emotional and cognitive responses, SRL processes, delayed trust–motivation mechanisms, and achievement outcomes. Reliability analyses showed that most constructs demonstrated acceptable to excellent internal consistency, although two epistemic belief subscales were excluded due to weak reliability. Multilevel mixed-effects models indicated that phase progression was the primary driver of achievement, with students improving significantly over time, particularly from Phase 1 to Phase 2. However, neither the main effect of group nor the Group × Phase interaction produced significant overall achievement differences, suggesting that the effects of AI dialogue support were not reflected in immediate performance gaps.

More informative patterns emerged in affective and regulatory trajectories. Group 3 showed relatively stable anxiety and protected positive affect, suggesting that structured dialogue may buffer uncertainty during problem solving. Group 4 experienced a temporary increase in anxiety and decline in positive affect in the middle phase, followed by recovery in the final phase, consistent with a productive-struggle interpretation in which increased epistemic responsibility initially imposes affective costs before supporting adaptation. In contrast, Group 5 exhibited less coherent anxiety trajectories, suggesting reliance on AI interaction without stable regulatory adjustment.

Motivational readiness strongly predicted SRL engagement, whereas human–machine trust showed a more complex role: higher trust was associated with weaker SRL engagement and lower achievement, implying that trust may sometimes reflect delegation to AI rather than epistemically productive collaboration. Bayesian multilevel mediation and delta-score robustness analyses did not support trust as a stable mediator of achievement gains, and epistemic belief maturity did not significantly moderate the trust–achievement relationship.

Overall, the findings suggest that the value of AI-supported dialogue lies less in generating immediate achievement advantages and more in reshaping trajectories of motivation, affect, and epistemic engagement. The study provides a process-sensitive account of AI-supported programming learning and highlights the importance of evaluating such interventions through longitudinal developmental adaptation rather than short-term performance differences.

Keywords:

AI-supported learning, epistemic agency, self-regulated learning, human–AI trust, programming education.