Learning-centred use of generative AI and later academic functioning: a baseline-adjusted three-wave panel study

Generative artificial intelligence (GenAI) is rapidly entering higher education, but measures of adoption, frequency, intention, or general attitude do not explain whether students use GenAI in ways that support academic work. This study focused on learning-centred use of GenAI in academic work (LCU), defined as a task-focused, verification-oriented, and non-substitutive pattern of GenAI integration, and tested time-ordered paths from LCU to later academic functioning through self-regulated learning (SRL) and academic self-efficacy (ASE).

We analysed a de-identified three-wave panel dataset from university students in China. The valid baseline panel included 1,200 students; 984 students were retained at T2, where SRL and ASE were measured in separated modules (T2a and T2b); and 788 complete cases were retained at T3, where academic procrastination and learning engagement were assessed. Baseline-adjusted composite-score path models controlled for prior SRL, ASE, procrastination, engagement, demographic characteristics, and GenAI-use covariates. Inverse probability weighting and auxiliary behavioural and rater indicators were used as sensitivity and convergence checks.

T1 LCU positively predicted T2a SRL (β = 0.345, p < 0.001) and T2b ASE (β = 0.118, p = 0.002), and T2a SRL positively predicted T2b ASE (β = 0.244, p < 0.001). SRL and ASE subsequently predicted lower T3 academic procrastination and higher T3 learning engagement after baseline outcomes were controlled. Indirect effects from LCU to both T3 outcomes were supported through SRL, ASE, and the SRL-to-ASE sequence; no residual direct LCU-to-outcome paths remained statistically significant after these mechanisms were included. Inverse probability weighting (IPW) and auxiliary-indicator checks showed a similar substantive pattern.

These findings show that the academic value of GenAI use is realised when students integrate it into learning-centred regulatory practice. Compared with frequency-based accounts, the present model identifies LCU as the behavioural condition that connects GenAI use with later academic functioning through strengthened SRL and ASE. The study advances a psychologically grounded account of GenAI in higher education and highlights the need to guide students toward task-focused, verification-oriented, and non-substitutive use that supports planning, monitoring, persistence, and engagement.