Students' engagement with ChatGPT feedback: implications for student feedback literacy in the context of generative artificial intelligence

Synthesis: A qualitative study of 16 undergraduates at a Hong Kong Professional Development university who used ChatGPT 3.5 to obtain feedback on IELTS writing tasks. Data came from unobtrusive screen-recorded observations plus stimulated-recall interviews. The study extends the traditional tripartite model of feedback engagement (cognitive, affective, behavioral) to a four-dimensional model adding metacognitive engagement, and asks what feedback literacy students need in a GenAI context.

Key findings by engagement dimension

Cognitive engagement — active, comparison-driven

• Three strategies: selective attention (n = 8, focusing on areas of self-assessed weakness), extracting key information (n = 4, filtering lengthy ChatGPT output), and comparison (n = 10 — comparing original vs revised versions, ChatGPT comments across time, against IELTS criteria, and against teachers' previous comments).
• Comparison is "a hub" for internal feedback (Nicol 2021): in GenAI contexts students can instantaneously compare multiple sources and versions — something hard in traditional feedback environments.

Metacognitive engagement — weaker, signaling "metacognitive laziness"

• Only 5 of 16 monitored their revision process; 6 reflected after finishing; none used goal-setting or planning strategies.
• Echoes Fan et al.'s (2025) metacognitive laziness: ChatGPT may boost short-term task performance while impeding self-regulation and knowledge transfer. Most revisions targeted local aspects of L2 writing (grammar, vocabulary, sentence structure), which may not transfer to future writing tasks.

Affective engagement — calm, but selectively trusting

• Majority (n = 13) described interactions as calm, relaxed, stress-free — minimal emotional resistance (unlike typical teacher-feedback contexts).
• Trust split: trusted ChatGPT on language forms (n = 11) but doubted its examples/evidence (n = 12); some doubted its legitimacy as an IELTS examiner.
• Negative affect arose when ChatGPT misunderstood prompts or repeated similar suggestions.

Behavioral engagement — superficial patterns

• 308 prompts total (9–40 per student); more prompts on local aspects (n = 104) than global (n = 55); 11.3% were direct copies of the draft; only 1.6% probed ChatGPT's credibility.
• Interactions were mostly one-off (n = 133 one-round vs 34 two-round vs 18 three-round).
• High uptake: 56.3% of ChatGPT comments honestly followed, 27.6% adapted, 16.1% rejected (83.9% overall uptake) — but revisions were superficial and local, and some students avoided directly using AI content due to academic-integrity beliefs.

Five capacities for student feedback literacy in a GenAI context

1. Prompt engineering — the quality of GenAI feedback is largely determined by prompt clarity (e.g. the CLEAR framework; Lo 2023).
2. Evaluative judgment — discerning useful feedback from plausible-but-unreliable output.
3. Emotional reflexivity — balancing trust and doubt by understanding GenAI's capabilities and limits (Bearman & Ajjawi 2023).
4. Ethical decision-making — deciding how, when, and why to use GenAI feedback so work remains authentic (academic integrity).
5. Metacognitive skills — setting feedback goals, planning prompts, self-monitoring interactions, and reflecting on the whole process.

What this means for practice

• Learners. Write prompts that name the target criterion — task response, coherence, lexical resource, grammar — instead of pasting the draft and accepting whatever comes back; prompt clarity largely determines the quality of what you get.
• Learners. Compare deliberately rather than absorbing: set ChatGPT's comments against your original text, your earlier drafts, the IELTS band descriptors, and your teacher's previous comments, and then decide what to adopt (uptake in this study was 83.9%, but the revisions stayed local and superficial).
• Learners. Set a feedback goal before you start and monitor your revision afterwards: only 5 of 16 participants monitored their process, 6 reflected only after finishing, and none used goal-setting or planning — the pattern the authors call metacognitive laziness.
• Learners. Push the revision beyond grammar and vocabulary: prompts focused on local aspects (n = 104) far outnumbered those on content and structure (n = 55), and local-only edits are the least likely to transfer to your next piece of writing.
• Learners. Verify before trusting: participants trusted ChatGPT on language forms (n = 11) but doubted its examples and evidence (n = 12), and only 1.6% of the 308 prompts questioned its credibility — and use the low-stakes calm of an AI exchange to ask the questions you would not bring to a tutor, while remembering that feedback literacy and engagement reinforce each other in both directions.

Limitations

• Sixteen undergraduates were recruited by convenience sampling from a single elective course at one Hong Kong teacher-education university, all using ChatGPT 3.5, so the four-dimensional model rests on one small, self-selected group.
• Each participant wrote one IELTS Writing Task 2 and revised it in a single sitting of 23 minutes to 1 hour 17 minutes; the study measures no later writing performance, so transfer claims are inferred rather than tested.
• No control or comparison condition with human feedback was used, and the trust, affect, and metacognition findings come from stimulated-recall interviews held a day after the task — self-report about one's own thinking, not directly observed mental process.
• None of the 16 participants had previously taken the IELTS test, so the revision behavior observed may not represent more experienced or higher-proficiency writers.

Citation

Zhan, Y., & Yan, Z. (2026). Students' engagement with ChatGPT feedback: Implications for student feedback literacy in the context of generative artificial intelligence. Assessment & Evaluation in Higher Education, 51(5), 821–834