Give Feedback at Scale Without Losing the Teacher

Feedback works when it arrives while the learner still cares about the work and is specific enough to act on. Both conditions fail at scale for entirely practical reasons: thirty pieces of work, thoughtfully commented, is several hours, and by the time it comes back the class has moved on.

This is a genuine constraint that AI can relieve, and it is also where the most consequential mistakes get made — because the same tool that drafts a helpful comment can quietly take over the judgement that should stay with a teacher.

Draft, judge, decide

The split that works is by kind of work rather than by volume.

Drafting is where AI genuinely helps. Identifying which rubric criteria a piece appears to address, pointing to specific passages, suggesting concrete next steps, and generating three different explanations of the same misconception are all real time savings, and they scale.

Judging should stay with the teacher. Whether this student's argument is actually weak or merely unfamiliar, whether an unusual approach is a mistake or an insight, and what this particular learner needs to hear next week are context-dependent judgements that the model has no access to.

Deciding — grades, progression, placement, any consequential determination — stays with the human, and this is not merely good practice. Automated decision-making about people carries regulatory obligations in many jurisdictions, and education is frequently treated as a high-stakes context.

The line is not about capability. It is that the teacher holds context the model does not have and accountability the model cannot carry.

Differentiate the support, not the expectation

Differentiation done well changes how a learner gets to the objective. Done badly it quietly lowers the objective for some learners, and AI makes the bad version very easy to produce at scale.

Legitimate axes: reading level of the material while the concept stays intact; scaffolding, more worked examples and more structured steps; modality, an explanation as a diagram, an analogy, or a dialogue; pace, more practice before moving on; interest, the same skill in a context the learner cares about.

The axis to protect: the objective itself. If one group is systematically asked to analyse and another to summarise, that is not differentiation. It is tracking, and over a term it compounds.

The check is simple and worth running: sort your differentiated versions by group and ask whether the cognitive demand is genuinely equivalent. It frequently is not, and nobody intended it.

Size the review gate to the stakes

Reviewing every AI-drafted comment eliminates the time saving. Reviewing none of it means shipping errors to students who may not be able to tell.

Scale the gate instead. Low stakes — formative practice comments — can go out with sampled review, checking a handful per set. Medium stakes — feedback on a piece that will be revised and graded later — should get a quick teacher pass on every item, which is still far faster than writing them. High stakes — anything affecting a grade or a record — needs full review, and the grade decision was never the model's in the first place.

Whatever the level, three things need checking specifically. Factual claims, because a confidently wrong correction is worse than no feedback and undermines the student's trust in all of it. Tone, because generated feedback tends toward a uniform encouraging register that can read as insincere, and can be harsh in ways nobody intended. And specificity, because generic comments that could apply to any piece of work teach nothing and students recognise them instantly.

Tell students when feedback was AI-drafted and teacher-reviewed. Discovering it later damages trust considerably more than knowing it up front, and disclosure also models the professional norm you want them to adopt.

Check who is actually benefiting

The equity question here is different from the access question, and it is easy to miss.

Students who already know how to use feedback — who read it, ask follow-ups, and revise — extract far more from an increase in feedback volume than students who do not. Adding volume can therefore widen a gap while looking like a universal improvement. Track engagement and revision behaviour by group, not just satisfaction.

The mitigation is to pair more feedback with teaching how to use it: a structured revision step, a short reflection on which comment mattered most, or a requirement to respond to one piece of feedback in writing.

Practice

Take one assignment set and run it three ways: your usual feedback, AI-drafted with sampled review, and AI-drafted with full review.

Measure your time per piece, and separately ask students which feedback they found most useful and whether they acted on it. Acting on it is the outcome that matters, and it is not always correlated with which feedback students preferred.

Then audit your differentiated materials for equivalent cognitive demand across groups. This takes twenty minutes and it is the check least likely to have been run.

Common mistakes

Letting drafting slide into deciding. Grades are not a drafting task, and in many places automated decisions carry legal obligations.

Differentiating the objective. Tracking with a friendlier name.

No review gate at all. A confidently wrong correction damages more than the missing comment would have.

Assuming more feedback helps everyone equally. It usually helps most the students who needed it least.