Research Article
Generative AI feedback use and English writing self-regulation among Chinese non-English majors: a two-wave study of feedback literacy and writing anxiety
Synthesis: Chinese non-English majors write a great deal of English for a course that gives them almost no individual Feedback, and chatbots have filled the gap. Yuan asked whether using that generative AI feedback relates to the self-regulation writing depends on, and whether the route runs through what students can do with feedback and how anxious writing leaves them feeling. Non-English majors at three Henan universities completed surveys twelve weeks apart: 520 of 650 were retained, with feedback use and feedback literacy measured first, and anxiety and writing self-regulation measured at follow-up; the chain was estimated with 5,000 bias-corrected bootstrap resamples. AI feedback use was positively associated with later self-regulation (c = 0.265), and 48.3% of that association ran through the two student-side processes, with the feedback-literacy route carrying the largest share. The design is observational and correlational, and the author says so repeatedly, framing every estimate as covariation rather than effect.
Key Findings
- Use predicted later self-regulation. Wave-1 generative AI feedback use was positively associated with wave-2 writing self-regulation (c = 0.265), and a direct association survived both mediators (c′ = 0.137).
- Nearly half the association ran through students. The total indirect effect was 0.128, or 48.3% of the total association, with a bias-corrected interval of 0.094 to 0.167 that excluded zero.
- Feedback literacy was the dominant route. The path through literacy alone contributed 0.086, 32.4% of the total association and roughly two-thirds of the indirect effect.
- Anxiety carried the smaller affective share. Lower writing anxiety accounted for 0.024, 9.1% of the total, while use kept a small direct negative association with later anxiety (a2 = −0.104).
- Competence came before relief. The serial route from use through literacy to lower anxiety to stronger regulation was positive and distinguishable from zero at 0.018, 6.8% of the total.
- Engagement, not volume, was measured. Restricting the predictor to the three feedback-seeking items kept the pattern (a1 = 0.310, indirect 0.109), so shared item wording did not produce it.
A feedback gap that chatbots filled
College English is required for nearly all non-English majors, taught in sections that commonly exceed fifty students, with one or two writing tasks per term, and drafts come back with scores and brief global comments rather than the draft-by-draft commentary writing development needs. Yuan recruited from that population and screened English majors out, since their curriculum supplies denser feedback. Wave 1 opened in April 2026, after about six weeks of instruction during which students had routine access to chatbots; wave 2 opened twelve weeks later. Of 650 valid wave-1 cases, 520 completed both waves (80.0%) across three universities, and completers did not differ from non-completers on feedback use, literacy, age or gender composition, so attrition gives no sign of having biased the sample.
Competence before emotion
The chain's second half links what students can do with feedback to how they feel about writing. Higher wave-1 feedback literacy was associated with lower wave-2 writing anxiety (d21 = −0.193), and lower anxiety was associated with stronger writing self-regulation (b2 = −0.231). Anxiety here is best read through control-value theory: it rises when an outcome matters but control over it feels doubtful, and literacy is a set of capacities that raise felt control over the revision cycle. Use also kept a small direct negative association with later anxiety (a2 = −0.104), consistent with private, non-evaluative feedback, though about half the literacy-anxiety path. On these coefficients, the judgment students make about a comment did more emotional work than access to it.
What the use measure rewarded
The predictor is an eight-item scale built for this study, covering three facets of a feedback episode: seeking AI commentary on a draft, reviewing the suggestions, and integrating them into revision. Its alpha was 0.883, and a single factor explained 55.1% of item variance. That wording matters, because students who paste a draft and accept the first output unread would score low. The author tested the overlap: three items ask students to check suggestions, approaching the judgment facet of feedback literacy, so the chain was re-estimated with only the three feedback-seeking items. The first path fell from 0.405 to 0.310 and the indirect effect to 0.109, and the serial pattern held. These are interaction patterns, not usage counts.
The association both mediators left
Even with literacy and anxiety in the equations, AI feedback use kept a significant direct association with later self-regulation (c′ = 0.137, about half the total). The author offers three readings. Students may imitate the rhetorical patterns they see in AI commentary, or gain task knowledge from the questions they ask; Self-Efficacy may profit from successful AI-assisted revisions without either mediator; and part of the residual may reflect the reliability limits of two short scales. Variance explained was 18.1% for literacy, 11.2% for anxiety and 19.9% for self-regulation, so most variance is unexplained. The two-wave design separates predictor from outcome but cannot rule out that capable self-regulators are simply the students who use AI feedback well.
What this means for practice
- Instructors. Teach the judgment, not just the access: the literacy route carried 0.086 of the 0.128 indirect association, so exercises in deciding which AI comments to accept are where the leverage sits.
- Instructors. Grade documented revision rather than the polished submission. The predictor rewarded students who checked suggestions before acting, and that behavior is visible in a revision log.
- Administrators. Blanket bans forfeit the literacy route while pushing use underground; embed norms where feedback is consumed, in drafting assignments and revision requirements.
- Researchers. Budget three waves with one construct per wave. Two waves leave both within-wave orderings resting on theory, which the author names as the design's central limit.
Limitations
- Two waves twelve weeks apart cannot separate within-person change from stable between-person differences, and the first link, use to literacy, is measured on the same occasion.
- No wave-1 baselines of writing anxiety or writing self-regulation were collected, so the R² values describe covariation with wave-1 predictors rather than change.
- All four constructs are self-reported; no revision logs, chat histories or text analyses were collected.
- The use scale is a preliminary, study-specific measure with no independent validation, test-retest evidence or criterion validation in other samples, and average variance extracted fell below 0.50 for all four constructs (0.294–0.487).
- The sample was a convenience group from three universities in one city, with three clusters too few for multilevel estimation and no measurement invariance tests.
Citation
Yuan, S. (2026). Generative AI feedback use and English writing self-regulation among Chinese non-English majors: a two-wave study of feedback literacy and writing anxiety. Frontiers in Psychology, 17, 2024870.