Bidirectional relations of epistemic beliefs and their impact on student engagement across GenAI platforms: a longitudinal quasi-experiment in middle school programming

Generative artificial intelligence (GenAI) is rapidly being adopted in educational settings, with many studies focusing on students’ learning behaviors and outcomes. However, little is known about how students’ epistemic beliefs (EB) develop over time and how these beliefs influence engagement in GenAI-supported environments. To examine the bidirectional relationships between EB and the impact of these beliefs on student engagement, a longitudinal quasi-experimental study was conducted with 395 seventh graders assigned at the class level to two GenAI learning environments: a domain-specific GenAI platform and a domain-general GenAI platform. Data were collected at the beginning of the semester (T1) and the end of the 13-week intervention (T2). Cross-lagged analysis indicated that EB related to source and complexity demonstrated strong temporal stability, with higher initial complexity predicting lower subsequent source and certainty. Multi-group path analysis further revealed that complexity and source positively influenced behavioral, emotional, cognitive, and agentic engagement, while certainty had a negative impact; these relationships were stable across platforms. These findings suggest that complexity drives changes in EB and that EB effects are robust across digital contexts, offering practical implications for fostering complexity beliefs, aligning GenAI tools with epistemic needs, and cultivating critical reasoning about GenAI outputs in classroom practice.