The two constructs at the heart of the study are relatively new to educational psychology. AI feedback literacy describes the competencies students deploy when confronted with feedback produced by an AI system: appreciating what such feedback can offer, judging whether it is trustworthy, managing their emotional reactions to it, and translating it into revised work. Agentic engagement, by contrast, is a behavioral construct referring to the proactive contributions students make to the flow of instruction, such as asking questions, expressing preferences, offering input, and suggesting alternative ways to proceed. Both constructs have been studied extensively in isolation, but the new research is presented as the first longitudinal test of how they interrelate within individual students over time.

The theoretical case for reciprocity drew on three lines of argument. Self-determination theory holds that students thrive when their needs for autonomy and competence are satisfied, and a student who can interpret and act on AI feedback is precisely the student for whom questioning the AI or contesting a suggestion feels rewarding rather than risky. Conversely, students who engage agentically accumulate the repeated, varied feedback encounters through which literacy is built. The study also invoked the mutualism principle, familiar from research showing that academic self-concept and achievement reciprocally fuel each other, and noted that AI feedback’s abundance and immediacy make the learner’s evaluative capacity, rather than access, the binding constraint on its productive use.

Methodologically, the random-intercept cross-lagged panel model matters because traditional cross-lagged panel designs conflate trait-like differences between students with state-like changes within them. The model partitions observed scores into a stable between-person component and fluctuating within-person residuals, then asks whether deviations from a student’s own typical level of one construct predict later deviations in the other, net of those trait differences. The six-week interval was chosen deliberately: shorter gaps risk capturing measurement noise before feedback-action cycles accumulate, while longer gaps would introduce confounds from midterm transitions and evolving familiarity with specific AI tools.

The sample comprised 263 women and 217 men, distributed across humanities, social sciences, natural sciences, and engineering. At baseline, participants reported using AI tools for academic purposes a median of three to four times per week. Of the 480 students who completed the first wave, 437 completed the second and 401 the third, and the 399 participants with complete data across all waves formed the analytic sample. Missing data were consistent with missing completely at random, and a priori power analysis indicated the design could detect cross-lagged effects as small as a standardized coefficient of 0.13.

The results supported both a stable between-person association and dynamic within-person reciprocity. At the trait level, students who scored systematically higher in AI feedback literacy across the semester also scored systematically higher in agentic engagement, with a moderate correlation between the two random intercepts of 0.315. At the within-person level, deviations in AI feedback literacy predicted later deviations in agentic engagement with a standardized coefficient of 0.222, while the reverse pathway was slightly larger at 0.263, both statistically significant with confidence intervals excluding zero. A Wald test found the two pathways did not differ significantly in magnitude, consistent with a symmetric mutualism in which each resource both predicts and is predicted by the other.

A model comparison underscored why the analytic choice mattered. The traditional cross-lagged panel model, which omits the random intercepts, fit substantially worse, and its cross-lagged estimates were markedly smaller, at 0.116 and 0.157, roughly 1.7 to 1.9 times smaller than the corresponding within-person estimates from the random-intercept model. This indicates that the trait-level covariation captured by the random intercepts was suppressing, rather than inflating, the apparent reciprocal dynamics in the traditional model. The authors argue this bidirectional bias means the traditional model should no longer be the default for reciprocity questions in feedback research, though the direction of its bias is not knowable in advance.

The authors are careful to bound their claims. Cross-lagged temporal precedence does not by itself establish causation: the design controls for stable between-person confounds and prior levels of each construct, but cannot rule out time-varying third variables, such as assignment difficulty or exam proximity, that simultaneously shift both constructs. Effects estimated as cross-lagged may also partly reflect contemporaneous influence within measurement occasions. Both constructs were measured by self-report, capturing perceived rather than enacted feedback use, though a marker-variable check and the model’s partialling of time-invariant method bias mitigated this concern. The sample, restricted to Chinese non-English majors, may not generalize to other systems, disciplines, or specific AI tools, and the rapidly changing AI landscape during data collection constitutes a further unmodeled confound.

The context of the sample is theoretically suggestive rather than merely limiting. Within Confucian Heritage Culture, agentic engagement has historically been characterized as culturally muted, with students socialized to defer to authoritative knowledge sources. The authors advance, explicitly as a hypothesis for future cross-cultural work rather than a conclusion supported by the data, the possibility that AI feedback may offer a less culturally loaded venue in which students can exercise agency, since there is no teacher to displease and no peer audience to read one’s questions as deficient. Whether the reciprocal pattern observed here merges with, or supplements, teacher-directed engagement in cultures where questioning authority is normalized remains an open empirical question.

Practically, the symmetric mutualism implies that interventions targeting either resource can, in principle, serve as an entry point into the developmental loop. But the authors caution against treating the two as fully interchangeable: because a core pillar of AI feedback literacy is judging when AI output is unreliable, encouraging open-ended agentic engagement before students possess baseline critical appraisal skills risks producing confident misuse, in which students interact energetically with AI feedback yet uncritically internalize its errors. They suggest educators may need to prioritize evaluative scaffolding early in students’ exposure, then let the reciprocal dynamic sustain itself. The broader reframing the study offers is that AI feedback literacy is not a static competence to be installed through training, but a developmental resource that grows through the very engagement it makes possible, a claim the authors note requires confirmation through longer time spans, experimental manipulation, and cross-cultural replication.

Subject of Research: Reciprocal longitudinal relationships between AI feedback literacy and agentic engagement in university students

Article Title: Within-person reciprocal processes between AI feedback literacy and agentic engagement: a three-wave random-intercept cross-lagged panel study

Article References: Xu, S. (2026). Within-person reciprocal processes between AI feedback literacy and agentic engagement: a three-wave random-intercept cross-lagged panel study. Frontiers in Psychology, 17, Article 1893874. https://doi.org/10.3389/fpsyg.2026.1893874

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1893874

Keywords: AI feedback literacy, agentic engagement, student engagement, feedback literacy, self-determination theory, mutualism principle, random-intercept cross-lagged panel model, longitudinal study, higher education, educational psychology, generative AI, learner agency