Generative AI feedback use and deep cognitive engagement in university EFL writing: a mixed-methods study

Synthesis: Feedback only changes writing when a learner does something with it, and He, Wu, Zhang and Chen (2026) ask what that something is when the feedback is machine-generated. They surveyed 486 university EFL students at three Chinese universities and followed up with 20 interviews, modeling reported GenAI feedback use as directly associated with deep cognitive engagement in writing and indirectly through learning autonomy and self-regulated learning. AI literacy strengthened the use–autonomy link, roughly doubling the slope at higher modeled literacy. Interviews show why averages mislead: a student with high measured literacy treated fluent output as fallible reference, one with low literacy used suggestions directly, and a third polished sentences under time pressure without asking why. Every construct is a same-time self-report, the instrument author-developed, and no writing outcome was measured.

Key Findings

1. AI feedback use went with deeper engagement. Reported use was directly associated with deep cognitive engagement even with autonomy and self-regulation in the model, and the model accounted for a bit over a third of the variance in engagement.
2. Autonomy and regulation added a second, indirect route. Use fed autonomy, autonomy fed self-regulated learning, and regulation fed engagement, producing a small but consistent indirect association.
3. Autonomy was feedback-specific control, not independence. Students described deciding when to ask, which paragraph to revise, and which suggestions to keep (S07, S12), rather than working without teachers or peers.
4. AI literacy strengthened the first link. The association between use and autonomy was roughly twice as steep for students at higher modeled literacy as for those at lower.
5. Interviews split on the authority of fluent output. The student with the highest measured literacy treated AI text as a fallible reference; the one with the lowest used suggestions directly because they seemed reliable.
6. Time pressure produced polishing without understanding. One student made a sentence "look better first" before thinking carefully about why, while another checked suggestions against the teacher's earlier comments despite limited AI familiarity.
7. Self-regulation and engagement were not empirically separable. The two self-reports moved closely together and no discriminant-validity test was run, so their distinction remains theoretical rather than demonstrated.

From using feedback to investing in it

The outcome the study cares about is cognitive rather than textual. Deep cognitive engagement is defined as elaboration, comparison of alternatives, reflective judgment and meaningful integration of feedback, explicitly separated from self-regulation, which concerns managing the activity rather than the depth of investment in it. That separation is what allows the paper to ask whether AI feedback is associated with more reflective revision rather than merely faster revision. The direct path is positive (β = 0.21) inside a model that already contains the regulatory route, and the full model accounted for 38% of the variance in engagement, which the authors read as evidence that frequent use alone is an incomplete account: S05's polishing shows an immediate purpose different from the evaluative work S03 and S15 described, where suggestions were compared with intended meaning and modified or rejected.

Autonomy is control over the feedback, not independence

Grounded in Self-Determination Theory, the model treats autonomy as perceived control over feedback access and revision decisions rather than as working without instruction. Reported use was associated with that control at 0.39, and control with regulation at 0.52 — the largest listed main-path coefficient, though differences between coefficients were never formally tested. Regulation in turn related to engagement at 0.47, and the serial product of those three paths came to 0.10 (95% CI [0.07, 0.14]). Interview accounts locate the perception in timing and access: S07 described asking the AI the moment they got stuck and feeling able to manage the writing alone, and S12 described choosing which paragraph to revise without waiting for a teacher's reply. The authors are careful about what this licenses. Convenient access can equally support avoidance, and their framing of autonomy as feedback-related agency does not establish that perceived control improved writing, reduced dependence, or was necessary for engagement. The distinction matters for teaching: a student can hold full control over when help arrives and still accept every suggestion unread.

AI literacy as a boundary condition

The moderation is the paper's most specific claim: the association between reported use and autonomy was stronger for students with higher modeled AI literacy (interaction β = 0.17; simple slopes 0.48 at one SD above the modeled mean and 0.23 at one SD below). The interviews give that interaction a face without verifying it. S09, whose AI-literacy mean was 4.25, judged both accuracy and fit with the text, saying "no matter how fluent the AI sounds, it can still be wrong, so I only use it as a reference"; S18, at 2.75, assigned the tool more authority than their own grammar, said "I generally use the AI's suggestions directly; they seem quite reliable," and rarely questioned errors. The authors offer the contrast as an interpretation, not as evidence that literacy causes better reliance. Two constraints on the construct matter for anyone reusing it: the four items measure perceived critical understanding, not performance-based competence, and they were written for this study, with one item measuring awareness that AI suggestions may be inaccurate. A further case cuts against a simple literacy story — S11 evaluated output against textbook examples and prior teacher feedback while describing their own AI knowledge as limited, which suggests feedback literacy can substitute for tool knowledge.

What a same-time self-report design can carry

All five constructs came from one questionnaire administered once, so the sequence in the figure is a theoretical ordering rather than an observed process. The authors are unusually explicit about the consequences: reverse and reciprocal paths remain plausible, alternative structural models were not fitted, and acceptable fit (CFI = 0.94, RMSEA = 0.057) cannot privilege this ordering over a statistically equivalent one. Shared-method influence is acknowledged rather than solved; Harman's single-factor test put the first unrotated factor at 32.4% of variance, which the authors themselves call insufficient. Measurement is the other soft spot. The 24 items were developed for this context, refined through expert review and a 48-student pilot, then validated on the same 486 respondents used for the structural model, with no independent sample. Self-regulation and engagement also shared item content closely (r = 0.57, against an approximate square root of AVE of 0.81), and neither HTMT nor a merged-factor model was estimated, which is why the key structural path rests on a distinction the data do not establish. Self-report also cannot distinguish perceived engagement from enacted revision.

What this means for practice

• Instructors: Require an explicit uptake decision rather than a polished draft: have students compare each AI suggestion with their intended meaning and with earlier teacher or peer comments, then record what they accepted, changed or rejected and why.
• Course designers: Ask for the reason separately from the rewrite. The student who under time pressure made a sentence look better could not explain the change, and that gap is the signal worth catching.
• Faculty developers: AI literacy is a plausible lever on whether feedback use is experienced as control, but the interaction is provisional and untested as training; evaluate any critical-evaluation module against writing quality and observed revision, not against satisfaction.
• Researchers: Pair repeated measures with interaction records and draft histories. Only designs that link what a student was shown, what they did with it, and how the text changed can separate perceived engagement from enacted revision.

Limitations

• The quantitative phase is cross-sectional: autonomous learners may seek more AI feedback, self-regulated learners may feel more control, and deeply engaged writers may request more explanations, so no path direction is identified.
• Same-source, same-time measurement may inflate the associations; Harman's test reports 32.4% on the first factor, and no marker variable or latent common-method-factor analysis was feasible.
• The instrument is author-developed (24 items, α = 0.85 to 0.91, AVE 0.607 to 0.654), refined through a 48-student pilot, then validated on the same 486 cases used for the structural model.
• Self-regulated learning and deep cognitive engagement share evaluative and reflective item content (r = 0.57, against an approximate square root of AVE of 0.81); neither HTMT nor a merged-factor model was estimated, so the key structural path rests on a distinction the data do not establish.
• The 20 interviews were purposively selected for variation, not for prevalence; only 10 were independently coded, that subset's rationale and a disagreement audit are undocumented, and the thematic work followed the quantitative results.
• Platforms and versions were not recorded — some interviewees mentioned ChatGPT, ERNIE Bot, Kimi or Tongyi Qianwen — and 32.5% of respondents had under six months of generative AI experience, so the findings concern unspecified tools rather than a particular system.

Citation

He, H., Wu, D., Zhang, B., & Chen, Z. (2026). Generative AI feedback use and deep cognitive engagement in university EFL writing: a mixed-methods study of autonomy, self-regulated learning, and AI literacy. Frontiers in Psychology, 17, 2006519.