Supporting self-regulated learning through generative AI feedback in online higher education: the importance of student perceptions of the source of feedback

Synthesis: Yilmaz and colleagues compared generative AI (GenAI) Feedback, informed by student trace data and learning analytics, against tutor-generated feedback in an online higher education statistics module, using a mixed-methods design with 46 blinded students. Students rated GenAI feedback more positively than tutor feedback, with a statistically significant edge on the Genuineness dimension, and the GenAI group showed significant improvement in the Task Strategies dimension of self-regulated learning (SRL). Qualitative insights revealed varied student awareness of the feedback source, with many adopting a content-first orientation while a subset indicated their attitudes might have shifted had they known the provider. The study concludes that GenAI can scale SRL feedback in online learning, but that student perceptions of the feedback source must be carefully managed for the expected impact to materialize.

Core Finding

Personalized, trace-informed GenAI feedback can match or exceed tutor feedback in perceived quality and measurably improve specific self-regulated learning behaviors (Task Strategies) in online higher education — yet its impact is conditional on how students perceive the source of the feedback, not merely its content. The significant gain in Genuineness under blind conditions, alongside a content-first attitude among many learners, shows that perceived value can outweigh concerns about origin when feedback is timely, tailored, and performance-specific.

GenAI Feedback vs. Tutor Feedback

Feedback is widely recognized as critical to learning, but providing timely, personalized feedback at scale remains a key challenge in large, diverse online cohorts where feedback gaps diminish Motivation and meaningful engagement. The study adapts Nazaretsky et al.'s (2024) framework of feedback perception — objectivity, usefulness, genuineness, and credibility of the provider. Under blind review, students rated GenAI feedback more favorably overall, with a statistically significant difference only on Genuineness, suggesting that the AI's consistent, supportive, and empathetic tone was interpreted as sincerity and care. The remaining dimensions (Objectivity, Usefulness, Credibility) showed higher GenAI means but no significant differences, echoing prior work (Henderson et al., 2025) that positions GenAI feedback as complementary to, not a replacement for, human feedback.

Self-Regulated Learning in Online Contexts

Online learning increases learner autonomy while reducing direct instructor presence, demanding advanced self-regulated learning skills. The study operationalizes SRL using the Ye and Pennisi (2022) evidence-centered framework, mapping roughly 48,000 MOODLE trace records to six proxies — goal setting, task strategies, environment structuring, time management, Help-Seeking, and self-evaluation. The GenAI treatment group showed significant improvement in Task Strategies, while the control group showed a lowered trend in Task Strategies and Time Management, consistent with prior findings that lower-SRL students are more prone to disengagement under inconsistent support. This demonstrates the value of trace-based indicators explicitly aligned to SRL constructs, rather than raw click counts.

Feedback Source Awareness

Students varied in whether they recognized the feedback as AI-generated, and awareness itself did not uniformly reduce favorability — a divergence from earlier disclosure studies. Many learners prioritized alignment with performance goals and practical value over origin, a pragmatic, content-first orientation; others indicated they would have scrutinized the feedback more critically had they known it was GenAI. These contrasting views reflect individual differences in academic proficiency, self-AI Regulation in Education level, and prior AI experience. The finding underscores that source perception is a socio-emotional determinant of feedback adoption that must be designed for, and it aligns with cautions about metacognitive laziness and cognitive offloading when AI support is used uncritically.

Relevance to the knowledge base

This paper contributes empirical, ecologically valid evidence on how Generative AI feedback shapes Self-Regulated Learning in Online Teaching and Learning, directly informing the knowledge base's understanding of AI Feedback Quality and the conditions under which AI feedback is adopted and acted upon. Its emphasis on Student-AI Interaction and source perception connects the technical promise of scaling feedback with the relational and socio-emotional realities of adoption. For practitioners, it argues that GenAI feedback should complement, not replace, human feedback, and that trace-based design can target specific regulatory behaviors while avoiding the pitfalls of over-reliance and metacognitive laziness.

What this means for practice

• Instructors. Keep human feedback in the loop rather than replacing it: this study does not claim GenAI feedback is superior to that of an experienced tutor, and the GenAI arm's only statistically significant advantage was on perceived Genuineness.
• Instructors. Ground AI feedback in trace data — the GenAI comments that scored well were built from personalized prompts sensitive to each student's performance and learning analytics records, not generic messages.
• Designers. Decide disclosure policy deliberately and support AI literacy: some students said they would have scrutinized the feedback more critically had they known it was GenAI, so source awareness shapes uptake beyond the feedback's content.
• Designers. Instrument specific regulatory behaviors rather than raw activity counts: the trace-based indicators aligned to six SRL proxies showed a significant Task Strategies gain, whereas click-count proxies in earlier work miss skill-level signals.

Limitations

• The sample is 46 students (23 per group) from a single online Basic Statistics module over nine weeks; the authors state this limits statistical power and generalizability and that the results should be read as context-specific.
• Only the Task Strategies dimension improved significantly; goal setting, time management, help-seeking, and self-evaluation did not, and none of the perception dimensions except Genuineness differed significantly — a pattern the authors tie to the small sample and to the granularity of the trace indicators.
• Outcomes are perception scores and behavioral traces rather than learning gains: the authors note that whether the feedback accurately and comprehensively addressed students' targeted needs was beyond the study's scope, and that the first author reviewed every GenAI comment before delivery — a workload that was never measured.
• The tutor knew the GenAI condition existed, which the authors concede may have prompted greater deliberation or altered usual feedback practice in unquantifiable ways.

Citation

Yilmaz, M., Temur, H. B., Emmungil, L., Çelik, E., Gauthier, A., & Cukurova, M. (2026). Supporting self-regulated learning through generative AI feedback in online higher education: the importance of student perceptions of the source of feedback. International Journal of Educational Technology in Higher Education.