When Generative AI Helps and When It Hurts Self-Regulated Learning: A Dual-Pathway PLS-SEM Model of Tanzanian Undergraduates

This study has many contributions Theoretically, it is the first empirical test of a dual-pathway model that simultaneously specifies the enhancement and offloading mechanisms of GenAI on SRL. Existing PLS-SEM studies in this area model only the positive pathway, which is inconsistent with the experimental evidence reported by Fan et al. (2024) and the qualitative evidence in the broader cognitive offloading literature. The dual-pathway specification produces falsifiable predictions about when GenAI helps and when it hurts. Methodologically, the study moves beyond the explanatory PLS-SEM tradition by incorporating predictive assessment (PLSpredict; Shmueli et al., 2019) and importance-performance map analysis (IPMA) on the focal outcomes. Predictive assessment is increasingly demanded by Q1 reviewers but remains uncommon in educational technology PLS-SEM applications. Contextually, the study addresses the geographic gap explicitly flagged by Xu et al. (2025) and Liu et al. (2025). It provides the first large-N empirical evidence on AI–SRL relationships from an East African undergraduate population, a context that differs from the Global North on infrastructure access, language of instruction, and prior digital exposure. The contextual contribution is not merely descriptive; the model permits formal multigroup comparison once comparable datasets become available from other regions. Practically, the IPMA results are designed to inform institutional policy at the participating universities and at comparable institutions across Sub-Saharan Africa. The analysis identifies which constructs combine high importance with low current performance that is, where intervention investment should be prioritised. This is more actionable than the generic recommendations that typically conclude correlational educational technology studies.