Generative artificial intelligence (GenAI) provides a new technical path for second language writing feedback. Compared with the traditional automatic writing evaluation system, which mainly relies on preset rules to identify language errors, generative artificial intelligence based on large language models can generate explanatory suggestions according to writing tasks and text content, and support learners in carrying out revision through multiple rounds of interaction. However, the ability to generate feedback is not equal to its teaching effect, which still needs to be tested in combination with specific writing results and the learning process.

Parra Núñez and Castrillo de Larreta-Azelain (2025) systematically analyzed 30 empirical studies on ChatGPT English writing feedback published from 2023 to 2025, and found that they played a positive role in language accuracy, text organization, and overall writing quality. The systematic review of Sulistiyo et al. (2026) further points out that the effective use of generative artificial intelligence feedback is related to dialogue interaction, multi-manuscript revision, and clear evaluation criteria; Students are usually more likely to adopt partial language error correction, while deep revision involving content and structure requires extra judgment and teaching support. This shows that the teaching value of artificial intelligence feedback involves not only the feedback information itself, but also how learners understand and deal with feedback.

The comparative study of different feedback sources shows inconsistent results. Ekizoğlu and Demir (2025) conducted an eight-week study on 60 senior high school English learners in Türkiye and found that the artificial intelligence-assisted feedback group was better than the teacher feedback group in improving the total writing score, grammar, vocabulary, and coherence. However, another quasi-experimental study on 60 college English learners found that the writing scores of the artificial intelligence feedback group and the teacher feedback group both improved, but the difference between the two groups did not reach a statistically significant level. These results suggest that it is not appropriate to presuppose that generative artificial intelligence feedback is necessarily superior to teacher feedback, but to examine its actual performance in combination with teaching arrangements, learners’ characteristics, and outcome indicators.

The measurement method of the feedback effect may affect the research conclusion. In the above study, pre-and post-writing tests were used to compare the teaching effect, but no delayed transfer test was conducted. Sulistiyo et al. (2026) also listed the long-term teaching effect as a problem to be further investigated. Therefore, in addition to immediate writing scores, it is helpful to further judge the learning effect associated with feedback activities by examining students’ performance in completing new tasks after removing tool support.

Learning transfer refers to the process by which learners apply the knowledge, skills, or strategies acquired in a specific situation to other situations. Perkins and Salomon (1992) distinguish between near transfer and far transfer according to the similarity between learning situations and application situations, and emphasize that transfer may not only depend on the automatic application of familiar skills in similar situations, but also require learners to consciously extract and apply abstract principles.

In English writing teaching, original revision and independent writing transfer belong to different evaluation levels. Learners can improve the current text according to external feedback, which can reflect their completion of specific revision tasks; The independent use of relevant language knowledge and text organization strategies in new writing tasks can provide more direct evidence of transfer. Especially in the case that generative artificial intelligence can directly provide suggestions for revision, it is difficult to distinguish the direct contribution of external tools from the development of learners’ own abilities if the learning effect is evaluated only according to the quality of the revised draft.

Therefore, this study defines English writing transfer as the performance of students applying the language knowledge and writing strategies obtained from previous feedback activities to new tasks without artificial intelligence and teachers’ immediate help. The study examines the performance of independent writing through immediate post-test and delay test, and measures the recent transfer and its retention by writing activities with similar task requirements and different themes. Considering that there may be a correlation between task difficulty and writing performance, the test questions need to go through a difficulty assessment and scoring consistency test.

Feedback literacy foregrounds learners’ active interpretation and use of feedback. Carless and Boud (2018) described four interrelated dimensions: appreciating feedback, making judgments, managing affect, and taking action. Their framework moves beyond a one-way transmission model and emphasizes learners’ agency in constructing the meaning of feedback and using it to improve performance.

In GenAI-feedback contexts, students must evaluate whether a suggestion is accurate, appropriate to the task, and consistent with their intended meaning. The distinction between local and substantive revisions identified by Sulistiyo et al. (2026) further motivates examination of feedback judgment and revision behavior. This study therefore focused on feedback understanding, judgment, and effective revision. A reasoned rejection of an inappropriate suggestion was treated as evidence of judgment rather than automatically classified as a failure to take up feedback.

SRL theory provides a broader account of learner agency. Zimmerman (2002) described a cyclical process comprising forethought, performance, and self-reflection. In English writing, these phases can be understood in terms of goal setting and strategic planning, monitoring during drafting and revision, and evaluating outcomes to adjust subsequent strategies.

Feedback literacy and SRL are related but analytically distinct. The former concerns how learners interpret, evaluate, and act on external feedback; the latter concerns how they plan, monitor, and regulate their learning in pursuit of goals. In the present study, students’ judgments and stated reasons for revisions informed the analysis of feedback literacy, whereas writing plans, monitoring practices, and reflective records informed the analysis of SRL. Both constructs may be associated with independent writing transfer, but the direction and strength of those associations require empirical examination and do not, by themselves, establish causal mediation.

Prior research provides evidence relevant to the use of GenAI feedback in English writing instruction, while indicating that findings vary across contexts and assessment measures. The quality of a revised text, immediate writing performance, and delayed independent writing performance capture different aspects of learning and should not be treated as interchangeable. The systematic literature reviews of Generative Artificial Intelligence Feedback have also shown closer attention to feedback uptake and longer-term outcomes.

From a theoretical perspective, feedback literacy concerns learners’ understanding, judgement, and use of feedback, whereas SRL concerns goal setting, monitoring, and reflection during writing. monitoring, and reflection in the writing process. Combining the two is helpful for analyzing the performance of independent writing from two aspects: feedback processing behavior and learning process adjustment, although the empirical associations remain to be established.

Based on this, this study takes junior two students as the subjects and adopts the non-equivalent group pre-test-post-test-delay calibration experimental design of two natural classes to compare the changes in students’ independent writing transfer and feedback processing behavior under the two teaching schemes of generative artificial intelligence feedback and teacher feedback, and to investigate the correlation between feedback literacy, self-regulated learning, and writing transfer. The main result of this study is the performance of independent writing without tools, combined with process evidence such as feedback judgment, composition revision, and learning reflection, to provide situational empirical reference for feedback teaching of English writing in junior middle schools.

A nonequivalent-group pretest-posttest-delayed-test quasi-experimental design was used. Participants were 82 Grade 8 students in two intact classes in City H. The experimental class (n = 41) received GenAI feedback, and the comparison class (n = 41) received teacher feedback. Both classes were taught by the same English teacher, used the same textbooks, and followed the same teaching schedule. Week 1 was devoted to the pretest, Weeks 2 - 7 to six weeks of writing-feedback instruction, Week 8 to the immediate posttest, and the delayed test to Weeks 3 - 4 after the immediate posttest.

To improve comparability, the study controlled for consistency among writing tasks, instructional content, class hours, revision opportunities, and assessment criteria. The study collected baseline writing scores and prior English achievement to describe initial differences. Because assignment was by intact class rather than randomization, class membership was confounded with feedback condition. Between-class comparisons are therefore exploratory analyses and do not support strict causal inference. Baseline writing scores, prior English achievement, and device access are presented in Table 1.

Table 1. Baseline writing performance, prior English achievement, and device access by class.

Baseline writing was scored out of 100 by averaging two raters’ assessments across task fulfillment, organization, vocabulary, grammatical accuracy, and appropriateness. Prior English achievement refers to the Grade 8 first-semester final examination (maximum 120). Continuous variables are presented as M ± SD and device-access variables as n (%). Independent-samples t tests and continuity-corrected chi-square tests were used. No statistically significant differences were detected on the baseline scores, English achievement, and devices, which does not establish equivalence on all learner characteristics or rule out unmeasured class differences.

The study comprised four stages: pretesting, instructional intervention, immediate post testing, and delayed testing (Table 2).

Table 2.Study procedure.

Both classes completed the same writing tasks following the sequence: independent drafting, feedback receipt, feedback evaluation, independent revision, and reflection.

The GenAI-feedback class used “seed 2.0 lite” to obtain writing feedback. A standardized prompt instructed the system: “You are a writing-feedback assistant for Grade 8 English learners. Based on four dimensions—task fulfillment, organization, vocabulary, and grammar—identify no more than three of the most important problems, explain why they are problematic, and offer suggestions that guide students to revise independently. Do not rewrite the entire text; acknowledge uncertainty when appropriate.” Under teacher supervision, students read the feedback, recorded reasons for their judgments, and revised independently. De-identified prompts and outputs were retained.

In the control class, the English teacher provided written feedback using the same assessment dimensions. Students completed feedback judgments, revisions, and reflective records under the same requirements as the experimental class. Both classes were taught by the same teacher and used identical writing tasks, assessment dimensions, and revision procedures. To improve comparability, a common upper limit was set on feedback items per composition, and expectations for explanations and classroom revision time were standardized. The AI class received feedback through a standardized prompt; the teacher-feedback class received written comments in the same dimension from the teacher. Differences in actual wording and explanatory content may nevertheless have remained, so the feedback source and its implementation are considered jointly as components of each instructional approach.

4.3.1. English Writing Transfer Assessments

Three independent writing assessments were administered: pretest (T0), immediate posttest (T1), and delayed transfer test (T2). No GenAI or other external feedback tools were permitted during any assessment in Appendix A.

Test tasks were designed according to Grade 8 English curriculum requirements and used the writing subject with comparable genres and different topics. Writing was scored on task fulfillment, organization, vocabulary, grammatical accuracy, and appropriateness, for a maximum of 100 points.

Two experienced English-language educators independently rated the essays without knowing students’ group membership. Interrater reliability was assessed using the intraclass correlation coefficient (ICC). Delayed-test performance was the primary outcome, while immediate posttest performance served as a supplementary outcome.

4.3.2. Feedback Literacy

The feedback literacy measure was informed by Carless and Boud’s (2018) student feedback literacy framework and adapted to junior secondary English writing. It assessed feedback comprehension, judgment, and use through 20 items on a five-point Likert scale, from strongly disagree (1) to strongly agree (5). Higher scores indicate higher self-reported feedback literacy. Full items, dimension assignments, and scoring rules appear in Appendix B.

A situated feedback-judgment task complemented the questionnaire. Whereas the questionnaire assessed students’ appreciation of feedback, confidence in evaluating suggestions, and tendency to use feedback, the situated task required students to assess proposed revisions and justify acceptance or rejection, thereby supplementing self-report measures.

4.3.3. Self-Regulated Learning

Drawing on Zimmerman’s (2002) cyclical model, writing-related self-regulated learning was operationalized in three dimensions: goal planning, process monitoring, and self-reflection. The questionnaire comprised 15 five-point Likert items; higher scores indicated higher self-reported writing self-regulation. The complete instrument and scoring rules are provided in Appendix C.

Students’ writing-reflection records collected during the intervention were also used as supplementary interpretive evidence.

4.3.4. Coding and Calculation of Process Indicators

Two process indicators were used. Because they differ in data source, denominator, and measurement occasion, their definitions are specified separately.

(1) Item-level effective revision rate. Data comprised all feedback items from three rounds of writing. Each student received three codable feedback items per round, yielding 369 items per class (41 students × 3 rounds × 3 items) and 738 items in total. Coding took place immediately after each revision round and concerned intervention-process data rather than test scores. Each item was coded for feedback accuracy, uptake, revision effectiveness, adequacy of justification, and justified rejection. A revision was effective only if it addressed the language or content issue identified, corrected or substantially improved the relevant passage, and introduced no new linguistic error or semantic distortion. Failure to meet any criterion meant that the revision was coded as ineffective. The effective revision rate was the number of effective revisions divided by all 369 feedback items in the class, regardless of uptake. If uptake items alone are used as the denominator, the measure is instead the effectiveness rate among adopted items: 50.4% (127/252) in the AI-feedback class and 41.2% (94/228) in the teacher-feedback class; these figures are reported only as supplementary information. Two coders independently assessed 20% of items (148 items), with agreement κ = 0.78; disagreements were resolved by discussion.

(2) Effective revision proportion in the situated task. At each administration, students evaluated 12 suggestions concerning the same essay, decided whether to adopt each, justified their decisions, and implemented revisions for adopted suggestions. The task was administered at pretest (Week 1) and posttest (Week 8). The effective revision proportion was the number of effective revisions divided by 12, the fixed number of suggestions.

These indicators should not be compared directly. The first uses all classroom feedback items as its denominator and reflects overall revision output during instruction; the second uses standardized suggestions and reflects judgment and revision performance under controlled task conditions.

Data included scores from the three independent writing assessments, feedback-literacy and SRL questionnaires, and writing revision records.

Baseline characteristics, writing scores across time, and questionnaire distributions were described, together with interrater and internal consistency reliability. An exploratory analysis of covariance (ANCOVA) then modeled delayed writing scores as the dependent variable and baseline writing scores as a covariate, reporting the adjusted between-class difference, effect size, and 95% confidence interval. Standard errors, significance tests, and confidence intervals from conventional ANCOVA assume, among other conditions, independent errors, whereas students within the same class may be correlated. Because each feedback condition involved only one intact class, treatment and class membership were perfectly confounded; class and feedback effects cannot be separated, nor can between-class sampling uncertainty be reliably estimated. Student-level P-values and confidence intervals are therefore descriptive of the fitted model and should not be interpreted as evidence of a generalizable intervention effect.

Correlations and parsimonious regressions examined associations of feedback literacy and self-regulated learning with independent writing transfer, with revision records used for interpretation. A sensitivity regression assessed whether the association between change in feedback literacy and delayed writing scores persisted after accounting for class differences and initial learner characteristics. In addition to baseline writing, changes in feedback literacy, and changes in self-regulated learning, the model included class membership (AI = 1; teacher feedback = 0), baseline feedback literacy, and baseline self-regulated learning. The stability and significance of the feedback literacy change coefficient were assessed. Given the sample size and intact-class design, no complex multiple-mediation analysis was conducted and associations were not interpreted causally.

The analyses included 82 Grade 8 students (41 per class). All students completed the pretest (T0), immediate posttest (T1), and delayed test (T2), with no missing observations. Writing scores, feedback-literacy scores, and SRL scores are summarized in Table 3.

Table 3.Descriptive statistics for the main variables by class.

Writing assessments were scored from 0 to 100; the feedback-literacy questionnaire from 20 to 100; and the SRL questionnaire from 15 to 75. Values are M ± SD; n = 41 per class, with no missing data.

Interrater agreement for delayed-test writing scores was ICC = 0.93 (two-way random-effects model, absolute agreement), indicating acceptable reliability. Cronbach’s alpha was 0.88 for the feedback-literacy questionnaire and 0.85 for the SRL questionnaire. Across the three dimensions of each questionnaire, alpha coefficients ranged from 0.76 to 0.81 and from 0.75 to 0.78, respectively. No items were identified as requiring removal or revision; the reported internal consistency supported subsequent analyses.

To address RQ1, writing performance was compared descriptively at baseline, immediately after the intervention, and at the delayed assessment.

In the GenAI-feedback class, the mean writing score increased from 61.02 (SD = 12.27) at T0 to 71.54 (SD = 14.76) at T1 and was 71.15 (SD = 14.03) at T2. Corresponding scores in the teacher-feedback class were 60.62 (SD = 12.67), 66.10 (SD = 16.26), and 64.74 (SD = 16.34). Both classes improved at T1, followed by small descriptive declines at T2 (0.39 and 1.35 points, respectively). Figure 1 depicts the trajectories.

In an exploratory ANCOVA with T2 as the dependent variable and baseline writing score as a covariate, the estimated adjusted between-class difference was 5.99 points, 95% CI [2.31, 9.67], F(1, 79) = 10.49, P = .002, partial η² = 0.117. The unadjusted difference was 6.40 points (d = 0.42, P = 0.061). A supplementary ANCOVA for T1 yielded an adjusted difference of 5.03 points, 95% CI [1.16, 8.89], F(1, 79) = 6.68, P = 0.012, partial η² = 0.078 (d = 0.35). The reported correlation between baseline and delayed writing scores was r = 0.82; adjustment for baseline increased statistical precision relative to the unadjusted comparison.

Table 4.Exploratory Comparisons of Independent Writing Performance.

Differences are defined as the GenAI-feedback class minus the teacher-feedback class; positive values indicate higher scores in the former. Models adjust for T0 writing scores; all F tests have df = (1, 79). The reported test of homogeneity of regression slopes was nonsignificant (P = 0.412). Unadjusted differences were 5.44 points at T1 (d = 0.35) and 6.40 points at T2 (d = 0.42).

The exploratory model showed an adjusted delayed-test advantage of 5.99 points for the AI-feedback class, 95% CI [2.31, 9.67], whereas the unadjusted independent-samples comparison was not statistically significant (P = 0.061). Because treatment condition and class membership were perfectly confounded, this model cannot identify the causal effect of feedback type, and its student-level P values cannot establish replicable class-level intervention effects (Table 4).

To address RQ2, the associations of feedback literacy and SRL with delayed independent writing performance were examined.

Pretest-to-posttest change scores were computed for feedback literacy and SRL and correlated with T2 writing performance (Table 5). Change in feedback literacy was moderately correlated with T2 scores (r = 0.41, P < 0.001), whereas change in SRL showed a weaker positive correlation (r = 0.26, P = 0.019). Baseline feedback literacy and baseline SRL were more strongly correlated with T2 performance (r = 0.63 and 0.55, respectively). Changes in situated feedback-judgment accuracy and effective-revision proportion were also correlated with T2 scores (r = 0.33, P = 0.003, and r = 0.20, P = 0.074, respectively).

Table 5.Correlations of feedback literacy and SRL with delayed transfer scores.

Change = posttest minus pretest. Pearson’s r is reported, with 95% CIs calculated using Fisher’s z transformation; n = 82. Correlations for changes in situated judgment accuracy (r = 0.33) and effective-revision proportion (r = 0.20) are not included in the table.

A parsimonious regression model included baseline writing score, change in feedback literacy, and change in SRL as predictors of T2 performance (Table 6). Variance inflation factors ranged from 1.02 to 1.35, indicating no marked multicollinearity among the predictors.

Table 6.Regression analysis of delayed writing transfer scores.

The dependent variable was T2 writing score. B = unstandardized regression coefficient; β = standardized coefficient. R² = 0.827; adjusted R² = 0.820; F(3, 78) = 124.0, P < 0.001.

Feedback-literacy change was positively associated with delayed transfer performance after adjustment for baseline writing. The manuscript reports an additional 15.0% of explained variance for feedback-literacy change (ΔR² = 0.150, P < 0.001). SRL change was weakly correlated with T2 performance in bivariate analysis (r = 0.26), but reportedly contributed only 0.7% additional explained variance after feedback-literacy change was included (ΔR² = 0.007, P = 0.070). H2 therefore received partial support: the feedback-literacy association was more robust, whereas the SRL association did not remain statistically significant in the multivariable model. No mediation pathway was tested.

A sensitivity regression examined whether these associations persisted after accounting for class and baseline learner characteristics (Table 7). After controlling for baseline writing, class membership, baseline feedback literacy, baseline self-regulated learning, and change in self-regulated learning, change in feedback literacy remained positively associated with delayed independent writing (B = 0.694, 95% CI [0.468, 0.920], P < 0.001). The class coefficient was not statistically significant (B = −0.109, P = 0.949), nor was change in self-regulated learning at the 0.05 threshold (B = 0.227, P = 0.0503). Adjusted R² was 0.817. The association between feedback literacy change and delayed writing remained after measured class and baseline differences were included; however, the two-class design precludes a causal interpretation.

Table 7.Multiple regression analysis of delayed writing transfer.

Note. Dependent variable: delayed transfer score (T2). B = unstandardized coefficient; β = standardized coefficient; 95% CIs are for B.

To contextualize the quantitative findings, all revision records were coded at the feedback-item level. A total of 738 feedback items were reported (three rounds per student, three items per round). Uptake and effective-revision rates were 68.3% and 34.4%, respectively, in the GenAI-feedback class, compared with 61.8% and 25.5% in the teacher-feedback class. The mean frequency of a “reasoned, appropriate rejection”—identifying an unsuitable suggestion and explaining why it should not be adopted—was 0.41 per student in the GenAI-feedback class and 0.32 in the teacher-feedback class. These descriptive process indicators suggest differences in both uptake and the quality of revision judgments.

The AI-feedback class had a higher baseline-adjusted delayed-test mean than the teacher-feedback class, with an unadjusted standardized difference of d = 0.42, a small-to-moderate effect. A difference was already apparent at immediate posttest (d = 0.35) and was somewhat larger at delayed test, suggesting that the observed pattern was not confined to the immediate writing assessment. One possible explanation is that reasons accompanying feedback helped students generalize individual corrections into rules they could later retrieve without tools. However, differences between classes may also reflect classroom climate, expectations, and other class-specific influences, and cannot be attributed to AI alone. Both classes showed some decline at delayed test, suggesting a need for periodic independent writing practice to reinforce learning.

The sensitivity regression further indicated that change in feedback literacy remained positively associated with delayed independent writing after controlling for class membership and initial writing, feedback literacy, and self-regulated learning. This suggests that the original association was not solely explained by average differences between the two classes or measured baseline characteristics. Nevertheless, because class and treatment were perfectly confounded and unmeasured factors could remain, the result supports only a relatively stable statistical association, not a causal or mediating role for feedback literacy.

Process data showed that situated-task judgment accuracy in the AI-feedback class increased from 52.0% to 77.0%, and the effective revision proportion from 45.5% to 76.6%, exceeding increases in the teacher-feedback class (52.8% to 60.2% and 42.1% to 53.5%, respectively). Both measures came from a situated judgment task administered at pretest (Week 1) and posttest (Week 8), each presenting 12 suggestions about the same essay. Students judged each suggestion and explained their decisions. The effective revision proportion used the 12 suggestions as its denominator, with the effectiveness criteria defined in Section 4.3.4. This measure differs from the item-level classroom effective revision rate in Section 5.3, which uses 369 classroom feedback items per class as its denominator. The former reflects judgment and revision under standardized conditions; the latter reflects overall revision output in ordinary instruction. For the teacher-feedback class, the classroom item-level effective revision rate was 25.5%, whereas situated-task effective revision proportions were 42.1% and 53.5% at pretest and posttest. Among 228 adopted classroom suggestions, 94 led to effective revisions, giving an effectiveness rate of 41.2% among adopted items. Because the tasks and denominators differ, the percentages are not directly comparable and do not establish equivalence of revision quality across measures.

Changes in situated judgment accuracy correlated more strongly with T2 (r = 0.33) than changes in self-regulated learning (r = 0.26), suggesting that evaluating and explaining feedback may be more closely related to transferable learning than mechanical uptake. Mean justified rejections were 0.41 and 0.32 per student in the two classes, a modest difference in the same direction. Some students could identify suggestions that sounded plausible but did not fit the text-behavior that uptake rates alone might incorrectly classify as insufficient engagement. Self-report and behavior did not always align: students reporting improved self-regulation did not necessarily show comparable improvement in revision behavior. Social desirability and self-report bias therefore warrant consideration.

Learning reflections indicated that students in the GenAI-feedback class more frequently reported “making a revision plan before starting” after the second round and began referring to rules summarized in previous rounds by the third round (e.g., “check tense against time expressions” and “each paragraph needs a topic sentence”). Such reports are compatible with a sequence of planning, monitoring, and reflection, but they are self-reported, and keeping a learning log may itself cue these behaviors. They are therefore treated as contextual process evidence rather than proof that transfer occurred.

Each feedback condition was represented by only one intact class. Classroom culture, teacher expectations, and peer interaction were therefore confounded with the feedback condition, limiting causal inference and generalizability. The sample may have been underpowered to detect smaller associations; imperfect equivalence of writing tasks, variability in AI output, and self-report bias are additional concerns. Future research should replicate the study across multiple classes and schools, ideally with cluster-level randomization and repeated follow-up assessments.