How Can AI Help Me Give Better Feedback at Scale?

It is week six. You have a stack of drafts, a rubric you believe in, and a growing suspicion that a third of your comments will be skimmed and forgotten. A colleague mentions that a tool now comments on every draft within the hour. You want to know whether that is a real upgrade or a faster way to produce text nobody reads.

The bottom line: AI feedback is worth adopting for coverage, timeliness and consistency inside a narrow band — copy-editing-scale language work and rubric-guided first-pass comments — and it is worth adopting only if you design what happens around the comments. Reach is the easy part. In a cluster-randomized experiment with 1,176 first-year undergraduates, reflective and hybrid designs beat direct AI critique on delayed, AI-free transfer; in a randomized essay experiment with 70 students, higher-quality AI feedback did not produce better revisions than an experienced teacher's. Across the evidence, what students do with a comment matters more than how fast the comment arrives, and feedback literacy — not AI access — separates the students who gain from the ones who copy.

The distinction that decides everything: volume is not quality, and delivered is not read

Two separations do most of the work on this page.

More feedback is not better feedback. AI can produce comments on every draft every week; the quantity is nearly free. Quality is not. Zhan, Boud, Dawson and Yan (2025) make prompt quality the hinge of the process: vague prompts yield generic, useless output, and generated comments can be hallucinated, biased or overlapping, which pushes evaluative judgment back onto the reader. Calibration is sharper still: in Farrokhnia et al. (2026), GenAI feedback quality was significantly associated with the strength of a student's initial draft, while teacher feedback quality was not — the teacher calibrated more consistently. Volume scales. Calibration does not.

Delivered feedback is not enacted feedback. This is the more expensive confusion. Farrokhnia et al. randomized 70 students writing argumentative essays in Persian into three groups. Chain-of-thought prompting produced significantly higher-rated feedback (M = 12.90) than both zero-shot prompting (M = 11.25, p = .01) and an experienced human teacher (M = 11.20, p = .008). The chain-of-thought group nevertheless did not revise its essays significantly more than the teacher-feedback group, which achieved comparable gains. Better-rated comments, the same revision. If your metric is the quality of what the model writes, that reads as a win; if your metric is what changed in the essay, it does not.

Who enacts feedback depends on the learner. In Hawkins, Taylor-Griffiths and Lodge (2026), 32 psychology students did a screen-recorded 25-minute essay task with unrestricted access to ChatGPT, then watched the recording back in a video-stimulated interview. Feedback literacy was the only significant positive predictor of essay grade (β = 0.46, p = .017); the special-issue editorial by Zhan, Wood, Carless and Yan (2026) notes that frequency of GenAI use, trustworthiness of the source and prior knowledge did not predict performance. Fewer than a third of the students compared AI output against another internet source; half expressed deliberate AI avoidance for academic integrity or wanting the essay in their own words; and most requested task-level feedback, which transfers poorly to other tasks.

Uptake is also unevenly distributed. In the university-wide Deakin pilot reported by Tubino and Adachi (2022), the AI tool was optional, and usage averaged roughly 13% of undergraduates and 12% of postgraduates despite almost 4,000 students having access. Proactive, high-achieving students used it more. That is what "making the tool available" buys you, and it is why the rest of this page is about design rather than access.

What AI is genuinely good at, and where it stops

Good: coverage and consistency inside a narrow band. Tubino and Adachi report that pilot at Deakin in 2021 using FeedbackFruits' AI automated feedback tool across 29 units and almost 4,000 students. It works on micro-level text features — sentence length, punctuation, grammar, text structure — supporting the copy-editing stage: the instructor sets parameters, the student uses the tool independently and gets timely, actionable feedback. The division of labor is what scales, not the model.

Bad: generality, correctness and calibration. That is where it fails (Zhan et al., 2025), and the remedy is a prompt, a rubric or a human — not a bigger model. Aldino et al. (2026) add instructor-side failure modes from 21 higher-education teachers: inconsistent tone (n = 7), potential misinformation (n = 5) and trust issues (n = 5), with revision effort spent deleting exaggerated praise and generic suggestions rather than adding content.

A deployment pattern that holds up

The evidence supports one shape more than any other: the AI drafts, the student evaluates first, and a human decides.

What to automate. Micro-level language work and rubric-guided first-pass diagnostics. Give the model a rubric and make it reason step by step — Farrokhnia et al.'s quality gain came from a chain-of-thought prompt walking the model through a step-by-step evaluation against an argumentation rubric, not a bare instruction. Gu, Chen and Yan (2026) likewise gave their GenAI group ChatGPT-4o with pre-trained rubrics and prompt guidelines, which is what made the output checkable against stated criteria. Rubric-referenced generation is automatable; judgment is not.

What to keep human. The final message, the tone, and anything that becomes a record. Becerra, Palma and Cobos (2026) describe AICoFE, which runs three independently fine-tuned models (GPT-4.1-mini, Gemini 2.5 Flash, Llama 3.1) over rubric scores and qualitative observations to produce independent draft comments, then has the instructor compose the final message by selecting sentences or paragraphs, with a legend showing which model contributed what. AI is a draft generator, not a final deliverer. Aldino et al. find the same "assist but verify" pattern: the ML component most often flagged Meeting Learning Objective as missing (20 of 21 teachers had omitted it; 16 accepted) and Student–Teacher Relationship (14 omitted; 12 accepted), with teachers deciding each suggestion on professional judgment.

How to sequence it. Do not stack sources side by side; stage them. In Ateş's (2026) experiment, the reflective condition required self-evaluation before AI critique, and the hybrid ran self-evaluation → peer feedback → GenAI critique. Tubino and Adachi propose templates for three drafting stages for the same reason. Combine AI with peer feedback rather than substituting one for the other: Gu et al. ran two parallel English classes (N = 118 first-year undergraduates in China; GenAI n = 56, peer n = 62) through three self-assessment cycles over one semester, and GenAI scaffolding improved feedback literacy slightly but significantly over human peer review (ANCOVA p = 0.049, ηp² = 0.03). GenAI students refined prompts iteratively and verified or challenged inaccurate output, while peer-group students chose sources by social convenience and had their evaluative judgment distorted by friendship bias. Peer review kept distinct value for audience awareness, so the authors propose multi-stage designs with anonymous peer feedback.

How to frame it for students. As a draft opinion to be judged, not an answer to be copied. Say plainly that model comments can be hallucinated, biased or overlapping and that the student owns the revision. The pair of IELTS writers in Zhan et al. shows the extremes: a low-literacy student with a vague prompt got generic feedback, trusted or over-copied it and engaged superficially; a high-literacy student wrote a criteria-referenced prompt, followed up, cross-checked sources and monitored revisions. The difference was instruction and habit, not tool access.

How to check the output. Sample generated comments against your own reading of the same drafts, looking for the failure modes named above — tone, misinformation, exaggerated praise, overlap. In the Deakin pilot, positive average ratings coexisted with student-flagged errors, so satisfaction is a weak quality signal; a smiley-face survey is not accuracy evidence.

Failure modes, with the numbers

Design, not availability, drives the effect. The strongest design evidence is a multisite, cluster-randomized, longitudinal field experiment: 1,176 first-year undergraduates in 48 sections across 4 universities and 3 science domains, randomized at section level into peer-only, direct GenAI, reflective GenAI and hybrid conditions. Hybrid produced the highest argument-quality gains and the clearest advantage on conceptual learning. Reflective and hybrid both produced stronger feedback uptake and self-regulated learning than direct GenAI, and both outperformed it on delayed, AI-free transfer. Direct GenAI did improve immediate argument quality over peer feedback — the advantage it has — but with weaker transfer. Educational value therefore depends less on AI access than on whether the feedback environment preserves student learner agency, evaluative judgment and ownership during revision: direct critique invites passive uptake, while reflective and hybrid designs force the student to interpret it, compare it against criteria, judge its relevance, and then revise. Other effect sizes temper this. Gu et al.'s advantage over peer review was small, and they read it as real but not transformative alone — teacher scaffolding, prompt guidelines and worksheets did the supporting work. The editorial's framing is that GenAI and human feedback are complementary only if the complementarity is specified, through sequencing, comparison, editing and governance rather than by leaving two sources side by side.

Over-trust at both ends of the pipeline. Students can over-rely uncritically when enacting feedback, and instructors can over-defer to suggestions. Teacher feedback is also often perceived by students as more negative or riskier than GenAI feedback, which inverts the usual quality assumption. Treat feedback literacy as a prerequisite rather than an assumption: the strongest predictor in Hawkins et al. was the learner's feedback literacy, not how often they used AI.

Workload that moves rather than shrinks. The workload question is the one most often promised away. Asked what the AI feedback tool gave them, 21 teachers named saving time (n = 2) only rarely, and named reflection (n = 14), improved language and structure (n = 11) and identification of missing components (n = 10) far more often. The visible gain is a draft and a diagnostic prompt, not a finished comment. What appears instead is review, calibration and verification work. Twelve of the 21 teachers made further sentence-level revisions; editing (f = 32) and removing (f = 27) dominated, with adding rare (f = 8). The dominant pattern was calibrating tone — editing praise down (f = 11), removing suggestions (f = 9), removing encouragement (f = 8) — to protect authenticity and professional voice, and revisions clustered in the relational dimension, especially Student–Teacher Relationship (f = 24). Teachers named the need for human editing (n = 9), inconsistent tone (n = 7), misinformation (n = 5) and trust issues (n = 5); instructors with more than five years of experience reported more of them, while less-experienced teachers valued the scaffolding — a reason to watch whether novices who defer to AI suggestions build less independent judgment. The editorial's rule: GenAI redistributes rather than removes teacher labor, and poorly designed tools can increase it.

Three objections, taken seriously

"AI feedback is not personal." Often true, and it is precisely the reason to keep the last word human. AICoFE exists because generic model prose is not what a student should receive — the instructor selects, edits and signs the final message. The 21 teachers in Aldino et al. rewrote tone for the same reason, deleting exaggerated praise and cutting suggestions that did not match their relationship with the student. Personalization is not the model's job in this design; it is yours, and the model is doing the setup work.

"My students will not read feedback anyway." They read less than we hope, and the evidence says the fix is design, not volume. Uptake in the Deakin pilot ran at roughly 13% of undergraduates and 12% of postgraduates. Farrokhnia et al.'s higher-rated AI comments produced no more revision than teacher comments. What moved the needle was sequencing that forces engagement — self-evaluation first, then peer feedback, then AI critique — plus three revision cycles across a semester that produced measurable feedback-literacy differences. Multiple same-day resubmissions are a usable trace of whether students engaged at all.

"This is not what students pay for." Then do not let the AI be the only voice they hear from you. The defensible version of the tool is the one where students get more feedback moments while you spend your hours on the judgment-intensive parts: the final word, the calibration, the comments that carry your authority. What students lose under careless deployment is not your presence but the coherence between criteria, comments and outcome. Specify the complementarity — sequencing, comparison, editing, governance — and the objection dissolves; leave two sources side by side and it does not.

This week

1. Sort your own last set of comments into language-level fixes and judgment-level calls. Only the first category is a candidate for automation.

2. Write the rubric-referenced, step-by-step prompt you would hand a model, and test it against three drafts you have already graded.

3. Sample the output against your own comments and log where the model was generic, wrong or flattering.

4. Stage one assignment as self-evaluation → peer feedback → AI critique, and tell students the AI is a draft opinion they are expected to challenge.

5. Ask for one comparison against a second source or against the criteria — fewer than a third of students do this unprompted.

6. Budget the review time you will actually spend on tone and deletion, and compare it with what you freed.

7. Watch the expert–novice split. If experienced colleagues report more problems with the tool's output than you do, that is information about your own calibration.

8. Decide what outcome counts and measure that: usage rates, revision traces and delayed unaided performance — not satisfaction ratings. The Deakin pilot's roughly 13% and 12% are a useful baseline, and AICoFE's curation tracking models the instructor side: log how much of the final feedback came from the model and how much you changed.

Where this page sits. Feedback is the umbrella concept for the whole system — provision, loop, uptake and assessment context — and AI feedback quality covers what makes generated comments accurate, useful, timely and pedagogically sound. This page is the practical middle: which combinations of those parts produce better feedback at scale. Questions about what a grade should mean when AI is involved — construct substitution, authenticated process evidence, integrity rules — belong to redesigning assessment in the AI era and assessment validity rather than here.

For the writing-instruction special case, see best practices for writing instruction in the context of AI.