Feedback Can Edit a Draft—or Build a Thinker
Generative-AI feedback in higher education raises a precise design question: is the system helping a learner judge and improve their work, or is it making the judgment on the learner’s behalf? This chapter examines a multisite experiment that compared direct, reflective, and hybrid feedback in scientific argumentation. The comparison shows how to choose a feedback process by its learning goal: rapid improvement of a draft is not the same as building judgment that survives when AI is unavailable.
The distinction matters because feedback is often treated as information delivered from one source to another. A comment can be accurate, detailed, and beautifully phrased yet teach little if the student neither understands it nor uses it. Educational feedback becomes consequential when the learner compares the work with a standard, decides what the comment means, and changes the work for a reason.
Feedback becomes learning through judgment
Three ideas make that process visible. Feedback uptake is what the learner does with critique: attends to it, interprets it, selects what is relevant, and implements a change. Evaluative judgment is the ability to decide about the quality of one’s own or another person’s work. Self-regulated learning is the broader cycle of planning, monitoring, evaluating, and adapting one’s cognition and behaviour in response to a task.
These are not abstract virtues added after the “real” work. They are the work. A student revising a scientific argument must decide whether a claim is supported by the evidence, whether the reasoning actually connects the two, whether an alternative explanation matters, and which change would make the explanation stronger. A system can point to a missing limitation, but the learner must still determine whether the limitation is real, how much it changes the conclusion, and how to revise without distorting the argument.
This responsibility is sometimes called epistemic agency: remaining answerable for what one accepts as adequate knowledge and reasoning. In an AI-mediated revision process, the key risk is not only that the system might be wrong. It is that the student might stop exercising the judgment needed to notice whether it is wrong. A fluent critique can become an answer key if the learner accepts it without comparison.
The study by Huseyin Ates tested this issue in introductory university science courses. It compared four conditions: peer feedback only; direct GenAI-supported feedback; reflective GenAI-supported feedback; and hybrid feedback that combined self-evaluation, peer feedback, and AI critique. Students worked on scientific argumentation, a useful test case because quality depends on more than clear wording. They must formulate claims, connect them to relevant and sufficient evidence, justify the connection, and address limitations or alternative explanations.
The four conditions differed chiefly in where evaluative work occurred and when it occurred. In the direct condition, students received AI-generated critique without first externalising their own diagnosis. This arrangement can be efficient: the system quickly identifies possible weaknesses and suggests revision moves. But it also places much of the initial evaluation in the system.
In the reflective condition, students first evaluated their own work against explicit criteria or identified uncertainties, then viewed the AI critique. That sequence creates a comparison. The student has already made a judgment that can be confirmed, challenged, or refined. The system is less like an authority handing down a verdict and more like an additional perspective that the learner must assess.