Modeling Student-Centered Feedback Mechanisms: ChatGPT-Supported Feedback in the Feedback Orientation Scale for Secondary Chinese Reading

Synthesis: Secondary Chinese reading classrooms in mainland China remain dominated by teacher-centered instruction, and access to transnational generative AI such as ChatGPT is restricted. Treating the Feedback Orientation Scale as a student-centered model of how external Feedback becomes learning, the study ran a two-month quasi-experiment with 200 seventh-graders (aged 12–13) from four intact classes in Shenzhen: 100 received ChatGPT-supported feedback and 100 conventional human-generated feedback. Multi-group structural equation modeling showed the pooled model fit poorly (RMSEA 0.259, SRMR 0.133), while the ChatGPT condition fit nearly perfectly (χ² 4.26, RMSEA 0.000, CFI 1.000) and the conventional condition only acceptably (RMSEA 0.112). The ChatGPT group gained 30.40 points on the feedback measure (dz 4.570) and 17.95 on feedback orientation (dz 2.552); the conventional group declined by −7.47 and −3.47. The claim is narrower than "AI feedback is student-centered": the mediating mechanism through feedback orientation held only in the ChatGPT condition.

Key Findings

1. The pooled FOS model failed; the ChatGPT condition's fit was excellent. Pooled fit was poor (χ² 71.83, RMSEA 0.259, SRMR 0.133), while Group 1 fitted closely (χ² 4.26, p = .642, RMSEA 0.000, CFI 1.000) and Group 2 acceptably (RMSEA 0.112, CFI 0.983).
2. Feedback orientation mediated learning only under ChatGPT support. The indirect FB → FO → WO effect was significant in Group 1 (β = 0.220, 95% CI [0.134, 0.317], p < .001) but not in Group 2 (β = 0.010, p = .141).
3. The ChatGPT group's gain on external feedback was very large. Feedback rose from 85.78 to 116.18 (mean difference 30.40, t(99) = 45.704, dz 4.570); Group 2 fell from 85.57 to 78.10 (mean difference −7.47, dz −1.297).
4. Feedback orientation moved in opposite directions. Group 1 rose from 29.28 to 47.23 (mean difference 17.95, t(99) = 25.523, dz 2.552); Group 2 dropped from 29.27 to 25.80 (mean difference −3.47, dz −1.337).
5. Reading outcomes improved in both conditions, but unequally. Group 1 moved from 24.87 to 42.54 (mean difference 17.67, dz 3.677), against Group 2's 24.80 to 26.42 (mean difference 1.62, dz 0.883).
6. Baselines were statistically indistinguishable. Pre-test ANOVAs were non-significant for feedback (F = 0.033, p = .856) and feedback orientation (F = 0.000, p = .989); all three post-test comparisons came in at p < .001 (df = (1, 198)).
7. The strongest pathways differed by condition. Group 1's FB_POST → FO_POST path was 0.845 and FO_POST → WO_POST 0.624; Group 2's fell to 0.273 and 0.111, with FB_POST → WO_POST at -0.053 (p = .270).

A student-centered model that fit only one condition

Zhu treats the Feedback Orientation Scale as the operational test of student-centeredness itself: if Feedback is genuinely personalized and responsive, the structure linking external feedback to outcomes through feedback orientation should hold. In the pooled sample it did not. That model produced χ² 71.83 (df = 5) with RMSEA 0.259 and SRMR 0.133, which signals structural misspecification. Splitting by condition resolved the failure in opposite directions. The ChatGPT-supported group fitted closely — χ² 4.26, p = .642, RMSEA 0.000, CFI 1.000, TLI 1.008, SRMR 0.018 — whereas the human-feedback group fitted acceptably but more loosely (CFI 0.983, TLI 0.967, RMSEA 0.112, SRMR 0.024). The reading offered is that the FOS is context-sensitive rather than a universal yardstick, and that teacher-centered conventional feedback maps onto it less cleanly than adaptive AI feedback does.

Feedback orientation as the mediating mechanism

The pathway evidence diverges most sharply here. Under ChatGPT support, prior feedback strongly predicted later feedback (0.782), later feedback strongly predicted feedback orientation (0.845), and feedback orientation in turn predicted reading achievement (0.624). Under conventional feedback, stability was high (0.859 for FO_PRE → FO_POST, 0.766 for FB_PRE → FB_POST) but the cross-variable links thinned: FB_POST → FO_POST was only 0.273, FO_POST → WO_POST 0.111 (p = .047), and FB_POST → WO_POST was -0.053 and non-significant. The conventional condition also carried a negative path from earlier feedback to later feedback orientation (-0.148, p = .040). Group 1's indirect effect (0.220) and total effect (0.313) were significant at p < .001, while Group 2's indirect effect (0.010) and total effect (-0.017) were not. Read against the self-regulated and metacognitive framing the author cites, the mechanism is mediation-dominant: feedback orientation carries the effect, not direct AI feedback quality.

Reading gains and what the conventional group lost

Both groups improved on reading outcomes, but unequally. ChatGPT-supported students averaged 42.50 (SD 4.18) at post-test against 26.40 (SD 4.41) for the conventional group, and the three post-test ANOVAs were decisive (feedback F = 774.161, feedback orientation F = 490.154, reading F = 703.039, all p < .001, df = (1, 198)). The declines are the more interesting result: the conventional group's feedback measure dropped to 78.10 and its feedback orientation to 25.80, both below pre-test levels (85.60 and 29.30). Because the study measured perceptions rather than only products, the pattern is consistent with students losing confidence in feedback they receive from a conventionally taught classroom. Zhu attributes it to routines that limit engagement and motivation in a context where ChatGPT-like tools are largely absent from prior secondary-level research.

Instrumentation and measurement quality

Three instruments anchor the results. External feedback was measured with a 31-item Student Feedback Perception Questionnaire on a 5-point Likert scale (31–155); feedback orientation with a 12-item Feedback Orientation Scale covering utility, accountability, social awareness and self-efficacy (12–60); and Chinese reading with a 50-point framework built on six comprehension processes — Retrieving, Explaining, Summarizing, Elaborating, Evaluating and Creating. Content validity came from three senior scholars in Chinese language education who reached unanimous agreement, and inter-rater reliability was strong (ICC .845 to .956). Measurement was nonetheless uneven over time: post-test SFPQ reliability was acceptable to good, but the pre-test was weak, with Cronbach's α of .505 and .413 for Dimension 2 in Groups 1 and 2 — a calibration problem the author attributes to baseline unfamiliarity.

What this means for practice

• Instructors. Integrate ChatGPT into feedback cycles to complement human feedback rather than replace it, and treat feedback orientation — whether students value, trust and act on feedback — as the thing to check, not just the score.
• Curriculum designers. Build reading tasks in which feedback arrives with proficiency-aligned texts and comprehension questions, then verify uptake, because here the gains followed the pathway through feedback orientation rather than the feedback itself.
• Researchers. Stop reporting pooled model fit for perception instruments across conditions; this dataset fit poorly when pooled (RMSEA 0.259) and closely when separated, so group-specific estimation should be the default.
• Administrators. Expect that a fixed conventional feedback routine can leave feedback-related measures below their starting point (−7.47 and −3.47); treat that as a program-level warning, not a null result.

Limitations

• 200 seventh-grade students (aged 12–13) from four intact classes at one public secondary school in Shenzhen were assigned by class rather than randomization, so causal claims rest on a quasi-experimental design.
• Group 1's Chinese language instructors were replaced by two computer science teachers offering only technical support, while Group 2 kept Chinese teachers with master's degrees and fewer than five years of teaching experience; teacher effects are confounded with the feedback condition.
• The intervention ran for two months (September 2025 to late 2025) with no follow-up measurement, so whether the feedback-orientation gains persist is untested.
• The pre-test SFPQ was psychometrically weak (Cronbach's α of .505 for Group 1 Dimension 2 and .413 for Group 2 Dimension 2), which limits how far baseline comparisons can be trusted.
• Feedback, feedback orientation and perceived feedback quality are self-reported on Likert scales, so the stalled pathway could reflect changed response habits rather than changed reading practice.

Citation

Zhu SiYi, A. (2026). Modeling Student-Centered Feedback Mechanisms: ChatGPT-Supported Feedback in the Feedback Orientation Scale for Secondary Chinese Reading. EdArXiv preprint.