AI Gives Corrective Feedback. My Students Still Didn’t Improve.

By Frank Carrizo Zirit

I recently set a writing task and permitted my students to use an AI chatbot for corrective feedback before submitting their final draft. The following week, I collected essays that had clearly been through an AI review. The sentences were cleaner. The vocabulary looked more sophisticated. But the arguments were still weak, the structure was still confused, and the same students who struggled to write well were still struggling.

That was my wake-up call. We had solved the problem of producing corrective feedback. We had not solved the problem of using it.

This article describes my mindset shift for corrective feedback in the age of AI. It is grounded in research on feedback literacy (Carless & Boud, 2018; Winstone et al., 2017), but most of all it is about what happened when I stopped focusing on comments and started focusing on interpretation. AI does not remove the need for feedback literacy. It makes it urgent.

The uncomfortable truth about feedback

We have long been trained to improve the quality of our corrective feedback: make it clearer, more specific, and timely. Yet the research tells us feedback is not powerful because it is delivered; it is powerful because it is used.

When students receive AI feedback, they often experience what I call the illusion of improvement. Surface-level textual changes create the appearance of development without corresponding growth in the writer’s conceptual understanding, rhetorical control, or task awareness. In AI-mediated contexts, this illusion is intensified because linguistic polish can be achieved independently of cognitive restructuring. The text improves; but does the thinker? Stevenson and Phakiti (2014) note that automated feedback can lead to modest improvements in draft quality, yet there is little evidence that this automatically transfers into deeper writing proficiency. Ranalli (2018) similarly highlights that students do not always know how to interpret automated corrective feedback, especially when it is generic or unclear.

The question is no longer Should we allow AI feedback? It is better to ask What kind of learner does an AI-rich classroom produce? In feedback research, this is often described as the problem of learner uptake: whether students actually process, internalise, and act on the information they receive (Winstone et al., 2017). If we do nothing, AI risks reducing uptake to surface edits and cosmetic changes. If we redesign our practice, it can strengthen genuine uptake, where feedback reshapes thinking rather than simply polishing sentences. In short – we can make sure learning happens.

The moment everything changed in my classroom

I banned the prompt “Rewrite this paragraph better.” My students were aghast; I soon learned that some had taken to asking AI to write their essays with mistakes baked in to make it look like student-produced work. Clearly, my students were more interested in getting a good grade than in improving their writing skills, and so I had to rethink assessment. Instead of only giving a grade for the final product, I decided to ask for proof of the work that got the student to that stage – much like how a maths teacher asks not only for the result but the working that went into the calculation. 

With that shift in place, I introduced three prompts the students were allowed to use: identify where my argument becomes unclear; explain why this paragraph lacks cohesion; and point out which sentences do not fully answer the task. This changed the student’s approach from outsourcing the task to diagnosing problems in their own writing. It was the beginning of a new rule in my classroom: AI can generate feedback, but you must generate judgment.

The R.E.A.C.T. routine: a structure that works

To make this shift sustainable, I introduced a structured routine called R.E.A.C.T., which operationalises the cognitive and metacognitive moves required to transform external feedback into internalised writing knowledge.

R – Recognise (Categorisation)
Students identify the type of feedback: grammar, cohesion, task response, tone, argument development. When feedback is unnamed, it feels overwhelming. When it is categorised, it becomes manageable.

E – Evaluate (Epistemic judgement)
Students ask whether the feedback is accurate, important, and aligned with the rubric. This is where critical thinking enters. Kasneci et al. (2023) remind us that large language models (LLMs) require new literacies; students must understand that confident tone is not proof of truth.

A – Articulate (Comprehension consolidation)
Students paraphrase the feedback in their own words. If they cannot explain it clearly, they have not understood it. 

C – Choose (Agency and prioritisation)
Students decide whether to accept, adapt or reject the feedback and justify their decision. 

T – Transform (Restructuring)
Students revise and then explain how the change improved the text. Not what they changed, but why it strengthened the writing.

Four practical changes that made the difference

If you are wondering what this looks like in practice, here are the adjustments that transformed my classroom.

1. The feedback log is compulsory

Every draft must be accompanied by a feedback log with four columns: feedback received; what it means; what I changed; and why I changed it. To make this concrete, here is the template I use:

Feedback received	What it means (in my words)	What I changed	Why I changed it
“Your argument lacks clarity in paragraph 2.”	My main point is not explicit enough.	I rewrote the topic sentence and added a clearer claim.	To ensure the reader understands my position immediately.
“Several sentences are too long.”	My ideas may be hard to follow.	I split two complex sentences into shorter ones.	To improve readability and cohesion.

At first, students complained. It felt like extra work. Then something interesting happened. Their revisions became deeper. They stopped fixing only grammar and started restructuring paragraphs.

2. I grade the thinking, not just the product

If we only grade the final draft, students will optimise for the final draft. That was the incentive structure I had unintentionally created. 

So I changed the incentive. I allocate marks for the quality of revision decisions. If a student thoughtfully rejects an AI suggestion and explains why it weakens their voice, they earn credit. If they prioritise structural feedback over minor grammar edits and justify that choice, they earn credit. Boud and Molloy (2013) argue that feedback must be embedded in assessment design; redesigning the incentive structure was my way of doing that.

Students quickly understood what counted. The message was clear: judgment, prioritisation, and reasoning were visible and assessable.

3. We analyse AI feedback together

Once a month, we all dissect some AI feedback together in class. We treat it as a text to analyse rather than a set of instructions to obey.

With mixed-level groups or lower-proficiency learners, I scaffold this process carefully. Instead of open discussion, we work through some simple binary questions, like Is the feedback clear? Is the feedback only about grammar? From there we can expand on the answer: “This feedback is important because…” or “I think this is unclear because…”. 

For slightly stronger groups, I use a Likert scale, asking them to rate the usefulness of each comment from 1 (not useful) to 5 (very useful) and explain their rating. The structure reduces cognitive load while preserving judgment.

4. We name feedback types explicitly

Many students treat feedback as one undifferentiated mass. I teach them to label it as surface correction, structural feedback, task alignment feedback or developmental feedback. This can be done with any tool available: printed drafts and highlighters, comment features in Google Docs, a shared Padlet board, an LMS discussion thread, or even sticky notes on the wall. Ranalli (2018) notes that automated feedback varies in usefulness across error types; once students understand categories, they prioritise more intelligently. Grammar moves from being the main event to being one part of a larger picture.

What changed in my classroom

After a term of this approach, three shifts became visible. Students asked better questions, moving from “Is this good?” to “Does this paragraph actually answer the task?” Weaker writers improved more than stronger writers because the structure gave them an entry point into revision. And the use of AI became calmer, less frantic, and less magical. AI stopped being a shortcut and started being a tool.

Why this matters now

Large language models are not going away (Kasneci et al., 2023). LLMs introduce a structural shift to the feedback ecosystem: they decouple revision from learning. Students can now achieve textual refinement without engaging in the evaluative processes that traditionally mediate development. If revision can occur without cognitive engagement, then learning becomes optional rather than necessary. Therefore, if we ignore AI, students will still use it, but if we forbid AI, they will use it quietly. However, if we redesign our classrooms around feedback literacy, something different happens. Students learn that feedback is not an answer key; it is raw material, and raw material requires processing.

A final thought

When I first allowed AI corrective feedback in my classroom, I thought I was modernising my practice. In reality, I had exposed a weakness in it. We had been teaching students to write; we had not been teaching them to revise with judgment. AI did not create that problem. It revealed it.

If you take one idea from this article, let it be this: do not ask whether AI gives good feedback. Ask whether your students know what to do with it. That question changed everything for me, and it might change your classroom too.

References

Boud, D., & Molloy, E. (2013). Rethinking models of feedback for learning: The challenge of design. Assessment & Evaluation in Higher Education, 38(6), 698–712.

Kasneci, E., Sessler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., et al. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, 102274.

Ranalli, J. (2018). Automated written corrective feedback: How well can students make use of it? Computer Assisted Language Learning.

Stevenson, M., & Phakiti, A. (2014). The effects of computer-generated feedback on the quality of writing. Assessing Writing, 19, 51–65.

Wineburg, S., & McGrew, S. (2017). Lateral reading: Reading less and learning more when evaluating digital information. Stanford History Education Group.

Winstone, N. E., Nash, R. A., Parker, M., & Rowntree, J. (2017). Supporting learners’ agentic engagement with feedback: A systematic review and a taxonomy of recipience processes. Educational Psychologist, 52(1), 17–37.

Winstone, N., & Carless, D. (2020). Designing effective feedback processes in higher education: A learning-focused approach. Routledge.

Biography