Making Sense of AI Feedback: Feedback Literacy Moderates ChatGPT Acceptance and SRL

Core Finding

Feedback literacy — specifically, students' ability to make sense of AI-generated feedback — is the boundary condition that determines whether ChatGPT acceptance translates into perceived self-regulated learning benefits.

In a survey of 211 Hong Kong secondary students (Grades 7–9), Mendoza, Xiong and Yan found that all five Technology Acceptance Model (TAM) components (perceived usefulness, perceived ease of use, attitude, intention to use, actual use) positively predicted self-reflection. But every one of these links was moderated by feedback sense-making: students with stronger feedback-processing skills reported greater self-regulatory benefits from ChatGPT use, while those with weaker skills showed minimal or even negative associations.

Design & Measures

Sample: 211 secondary students (Grades 7–9, ages 12–15; 51% female; 102 G7, 49 G8, 60 G9) from a government-funded Hong Kong school running AI literacy initiatives. Data collected October 2023 via online survey; nine of 220 initial respondents excluded for suspicious patterns/outliers.
Acceptance: 25-item ChatGPT Technology Acceptance instrument (adapted from Davis, 1989; reliabilities .90–.95), measuring Perceived Usefulness (PU), Perceived Ease of Use (PEOU), Attitude (ATT), Intention to Use (ITU), and Actual Use (AU).
Moderator: 4-item Making Sense of Feedback subscale from Dawson et al.'s (2024) Feedback Literacy Behaviour Scale, adapted to reference ChatGPT (α=.94) — capturing credibility assessment, decision to use, judgment of conflicting comments, and standards alignment.
Outcome: 7-item self-reflection subscale from Yan's (2018) Self-assessment Practice Scale (α=.93).
Analysis: Hierarchical multiple regression (demographics controls → main effects → interaction), Johnson-Neyman technique, and simple-slope analyses at low/moderate/high sense-making. >90% power for medium effects; VIFs <3.5.

Results

Main effects (H1 — supported)

All five TAM components significantly predicted self-reflection after controlling for age, gender, and socioeconomic status:

Component	β	p
Perceived Ease of Use	.45	<.001
Perceived Usefulness	.40	<.001
Attitude	.39	<.001
Intention to Use	.30	<.001
Actual Use	.27	<.001

Moderation (H2 — supported)

Feedback sense-making significantly moderated each link (interactions β=.11–.20, all p<.05; full models R²=.28–.34; ΔR² from step 2 = .124–.234).

PU & PEOU: significant associations with self-reflection emerged only at moderate-to-high sense-making (Johnson-Neyman thresholds 3.89 and 3.55). Slopes were non-significant at low sense-making.
Attitude: the positive relationship appeared only at high sense-making (above 4.26).
Intention to use: a crossover interaction — negative at low sense-making (β=-.17, p=.01) but positive at high (β=.15, p=.02). Students with weak evaluation skills but strong intentions to use ChatGPT may be exhibiting over-reliance rather than strategic use.
Actual use: the interaction was significant but only for extreme-low sense-making (below 2.00, outside the observed range); within typical student capabilities there was no reliable conditional relationship.

Interpretation & Implications

The authors position feedback sense-making as a fundamental cognitive prerequisite for productive AI engagement, arguing the "black box" nature of generative AI places unique evaluation demands on students. The crossover pattern for intention to use supports a distinction between problematic dependency and strategic resource utilisation.

Practically, they recommend a three-tier implementation framework (universal instruction in evaluating AI credibility → scaffolded interaction for moderate-literacy students → independent use for high-literacy students) and propose a 4.2 threshold on the Feedback Sense-Making subscale to identify which students report benefiting from ChatGPT acceptance.

Methodological Strengths & Caveats

Strengths: strong reliability across all scales, power analysis for regression, assumption checks, and Johnson-Neyman analyses that map where effects turn significant.
Caveats: cross-sectional design (no causal inference), self-report measures rather than observed behaviours or objective outcomes, single educational system, and data collected during the early-positive-coverage adoption period of ChatGPT.

Connected Concepts

Connected Articles

Citation

Mendoza, N. B., Xiong, Y., & Yan, Z. (2026). Making sense of AI feedback: how students' feedback literacy moderates the link between ChatGPT acceptance and self-regulated learning. Educational Psychology. DOI: 10.1080/01443410.2026.2641521. CC BY.