GenAI as a Learning Partner: Supporting Self-Regulated Learning Over Time Without Replacing Effort

Synthesis: GenAI as a learning partner — a longitudinal study of 75 first-year university students across a full semester showing that SRL is both a stable aptitude and a dynamically fluctuating state. Individual baselines were consistent, but metacognitive knowledge and Well-Being declined systemically over the term, driven by curriculum demands (e.g., major assessment deadlines). A proof-of-concept demonstrated that giving an Large Language Models (LLMs) personal, temporal, and contextual information enables it to identify tailored SRL support directions.

Key Findings

• SRL as both stable and dynamic. Pre-/post-semester surveys showed high individual consistency (R² up to 0.73), but paired t-tests revealed systemic declines in metacognitive knowledge (MAI_K) and well-being (WEMWBS) over the term — while metacognitive AI Regulation in Education (MAI_R) held steady.
• Curriculum drives fluctuations. Week-by-week daily-survey analysis showed a major Week 8 project deadline lowered interest, Self-Efficacy, and emotional well-being, with most affective metrics recovering after the assessment. Students shifted from planning/reviewing toward reading new material after the deadline.
• Context-aware GenAI support. By separating formal, psychometrically-validated measurement of SRL from the LLM's interpretation, the authors gave Gemini 3 each student's baseline, week-by-week fluctuations, and academic context — and it generated appropriately tailored support (e.g., foundational strategy-building for a low-metacognition student vs. post-assessment recovery pacing for a high-metacognition student whose confidence had dipped). Researchers judged the output pedagogically sound.

What this means for practice

• Instructors. Map AI support to the academic calendar rather than the student's baseline profile: reduce unsolicited help during deadline weeks so students keep the productive struggle, and raise support after assessments to rebuild self-efficacy and interest.
• Instructors. Keep measurement and interpretation separate — score SRL with validated instruments (MAI_K, MAI_R, WEMWBS) and let the model only translate those results into support directions, so advice stays anchored in theory rather than inferred from chat history.
• Instructional designers. Feed the system each learner's personal, temporal, and contextual data (baseline, week-by-week fluctuations, course context) instead of a one-time-fits-all personalization built only on baseline aptitude.
• Instructional designers. Design for SRL as a state, not a trait: instrument Metacognition and well-being across the whole term, since both declined systemically while metacognitive regulation held steady.

Limitations

• Although 75 students consented, only 54 completed both pre- and post-semester surveys, and the week-by-week analysis narrowed further to a 38-student cohort in a single computing subject at one Australian university.
• Every measure is self-report (MAI_K, MAI_R, WEMWBS, and single-item interest and self-efficacy scales), so the reported declines reflect perceived rather than demonstrated self-regulation or learning.
• Daily surveys ran only in weeks 4, 6, 8, and 9 (246 responses), so the "week-by-week" trajectory rests on four sampled weeks of the 12-week term.
• The GenAI component is a proof-of-concept on two illustrative student cases whose output was judged appropriate by two members of the research team; there was no control group and no measured effect of the tailored support on learning.

Citation

Song, Y., de Barba, P., & Oliveira, E. (2026). GenAI as a Learning Partner: Supporting Self-Regulated Learning Over Time Without Replacing Effort. Learning Letters, 8, Article 73.