AI-Assisted Marking: What to Automate, What to Judge Yourself

Feedback works best when it is fast — a correction that arrives while the attempt is still fresh is worth far more than the same correction a week later. Marking, meanwhile, is slow, and it consumes more teacher time than almost any other task. AI-assisted marking sits exactly on this tension, and the honest answer to whether it works is: it depends entirely on what is being marked.

Reliability varies enormously by task

At one end, objective formats — multiple choice, numeric answers, matching — are marked perfectly by simple automation, no AI judgement required. Short factual answers are next: modern systems handle rephrasings well, though they occasionally accept a confident wrong answer or reject an unusual correct one. Then the gradient steepens. Marking a student's method — awarding partial credit for correct working with one slip — requires exactly the step-by-step reasoning where AI is least dependable. And extended writing sits at the far end: AI can assess surface features of an essay competently, but genuine evaluation of argument quality, originality, and voice remains unreliable, and its scores can be swayed by fluency over substance.

Where teacher judgement is irreplaceable

• Partial credit for method: deciding that a student understood the concept but slipped on arithmetic is a judgement about understanding, not about text.
• The diagnosis behind the error: two students with the same wrong answer may need entirely different help, and telling them apart requires knowing them.
• Anything with stakes: grades that affect streaming, promotion, or reports should never rest on an unreviewed automated score.
• Sensitive content: a personal essay about a difficult home situation needs a human reader, full stop.
• The relationship: students work for teachers who they believe actually read their work. Fully outsourced marking is noticed, and it corrodes effort.

AI feedback as first-draft feedback

The most defensible model for written work treats AI output the way an editor treats a junior colleague's notes: a first draft of feedback, reviewed and owned by the teacher before it reaches the student. The AI reads a class set of paragraphs and drafts a comment on each; the teacher skims every one, deletes the wrong ones, sharpens the vague ones, and adds the line that only someone who knows the student could write. Teachers using this pattern report that reviewing thirty drafted comments is dramatically faster than composing thirty from scratch — and the student still receives feedback a person has stood behind.

Two failure modes deserve naming. First, rubber-stamping: under time pressure, review quietly degrades into approval, and errors start flowing through to students with your name attached. Decide in advance what percentage of comments you expect to edit — if you are editing almost none, you are probably not reviewing. Second, generic praise: AI feedback drifts toward pleasant, interchangeable comments that could attach to any essay. Feedback that does not name something specific in this student's work is decoration.

A sensible division in practice

1. Automate the objective layer fully — quizzes, drills, numeric homework — and spend none of your attention there.
2. Use AI-drafted comments, human-reviewed, for routine formative written work.
3. Mark summative and high-stakes work yourself, perhaps with AI as a second opinion after your own judgement is formed.
4. Reinvest the saved hours where they compound: reviewing error patterns across the class, and the ten minutes of lesson time where students act on the feedback.

That last point is the one worth ending on. Faster marking is only valuable if the speed buys something. The research on feedback is blunt: it works when students respond to it, and much of it is otherwise wasted. AI can compress the loop from a week to a minute — but only a teacher can make the student close it.