AI Marking and How Fast Feedback Has to Be Before It Counts

Friday recess, and thirty-two S2 English journals are stacked on the corner of the staff room desk. The teacher puts a hand flat on the pile: "I'll mark these tonight, they get them back Wednesday, and by then not one of them will remember what they wrote." Two evenings of work, give or take.

The problem with marking is not the number of hours. It is that by the time the comment reaches the student, the piece of work has stopped mattering to them. Cutting the wait from three days to three minutes changes willingness to revise entirely.

The shelf life of feedback

The same sentence has very different force depending on when it lands. Right after writing, a student still remembers getting stuck in the second paragraph and not knowing how to carry on. Three days later, all they see is a grade and some red circles, and the place where they got stuck is gone.

The window where students revise

Instant feedback is valuable because it arrives while the student still cares about the piece. That S2 class tried spending the last eight minutes of a lesson typing their journals into the platform, getting paragraph-level feedback on the spot, and revising once before submitting. Twenty-six of the thirty-two actually rewrote at least one paragraph. The identical comments, appearing on paper three days later, normally move fewer than five.

Fast does not mean less marking

Instant feedback does not replace the teacher's pass. It answers "where did I go wrong, right now". The teacher's pass answers "which single weakness has this student not moved on all term". The second question is not one anyone else can answer for you.

How fine a rubric has to be

Roughly eighty per cent of marking quality is decided by the rubric. Panels tend to write bands like "substantial content, vivid description, complete structure", which helps the model exactly as much as it helps a thirteen-year-old: not at all. Nobody knows how much "vivid" is.

Rewrite bands as visible actions and the picture changes immediately. For the content dimension of a narrative, "the event has a time, a place and at least three specific actions, and emotion is conveyed through action or dialogue in at least two places" gives the model something to check item by item — and gives the teacher a short list to verify during review.

This is panel work, not individual work. Once the rubric is fixed, the whole subject marks with one ruler and comparing across classes starts to mean something.

Three kinds of questions, three levels of reliability

Treating all marking as one activity is where most of the disappointment comes from. These three are not close.

Objective items and worked steps

Multiple choice, fill-in-the-blank and computation are the stable ground: given a clear answer key, errors are located precisely. Worked steps go further, marking not just the final answer but the point where the work went off. On a linear equation, the feedback reads: "Step 3 — you moved 5x across the equals sign without changing its sign, so everything after that inherits the error." The student sees a break point rather than a cross.

Open writing

Here the output is a first pass, nothing more. The model is genuinely reliable on the language layer — inconsistent tense, three consecutive sentences opening with "And", one word appearing seven times on a page. On judging what a piece is worth, it is lenient and it crowds scripts into the upper-middle bands. That is precisely why the teacher's pass cannot be dropped.

On open writing, ask the model to locate problems, not to decide grades.

The three gates teachers keep

The first is marking three by hand before anything else. On each new assignment, mark three scripts yourself, then look at what the model made of the same three. A wide gap usually indicts the rubric, not the students.

The second is sampling after the fact. Once the batch is done, pull five: the top script, the two lowest, and two from the middle at random. The two lowest matter most — misjudging a weak student costs far more than misjudging a strong one.

The third is who owns the mark. No AI grade should appear to a student as official before a teacher confirms it. That one is a matter of principle, and it belongs in the school's written guidance rather than in a verbal reminder.

What the student sees in red

Accurate feedback that no student reads is worth zero. Three requirements are worth enforcing.

Quote the script. "Watch your structure" cannot be acted on. "In paragraph four you wrote 'I was so scared' — try writing what you actually did at that moment" can.

Flag, don't fix. Have the model mark the problem and name its type, but withhold the corrected sentence. The moment a corrected version appears, it gets copied and the revision never happens. Decide this setting before the assignment goes out, not after.

No more than three points at a time. A script returned with every error marked simply makes a weaker student give up. Choose the three that most obstruct meaning and leave the rest for the next piece. That is what scaffolding means in practice — hold up the parts that matter most, and let go by degrees once the student is steady.

Common misreadings

• The claim: AI marking hands the red pen to a machine and the teacher stops marking. What actually happens: What disappears is the time spent flagging the same error thirty times. Judgement time stays. The teacher's pass shifts from hunting typos script by script to reading which dimension the whole class failed — different work, not less.

• The claim: The more detailed the rubric, the more accurate the marking. What actually happens: Neither a model nor a teacher can hold twelve dimensions steady, so every one of them gets marked thinly. Four or five dimensions, each band written as a concrete action, beats twelve every time.

• The claim: The grades are unreliable, so the whole thing is unusable. What actually happens: Grades and feedback are separable. Even when the teacher regrades from scratch, the paragraph-level problem list has already saved the student three days of waiting — and those three days were when they were most willing to revise.

Three things to try on Monday

1. Take a rubric you currently use, find its vaguest band descriptor, and rewrite it as something you could tick off — "vivid description" becomes "conveys emotion through action or dialogue in at least two places".
2. Pick an assignment you have already marked, run three scripts through the model, and compare bands. Whichever dimension shows the widest gap is the next part of the rubric to fix.
3. Before the next assignment goes out, say one sentence to the class: instant feedback is the first reader's opinion, the teacher decides the grade, and revised work can be resubmitted.