Second marking with AI: a documented second read

Second marking explained — double, blind and check marking — and how an AI second read can sit inside your policy, with agreement data visible to QA.

Second marking is the practice of having a second marker review work that a first marker has already marked, against the same criteria, so that no result rests on a single judgement. Most institutions already run a second-marking policy, and the question a QA office asks about AI is narrower than whether AI should mark: can an AI second read sit inside that policy without weakening it? NUADU's claim is deliberately small. AI can act as a documented second rater within your existing policy, with agreement, override and drift data visible to QA.

What is second marking?

Second marking is a quality-assurance step: a second marker reads work that has already been marked, against the same rubric, and the two judgements are compared before a result is confirmed. The second marker is usually a colleague with the same subject expertise, sometimes a moderator from outside the module team. Policies usually combine several forms of second marking, chosen by stakes and volume.

  • Double marking — two markers each reach their own judgement, and any difference is resolved through an agreed process.
  • Blind double marking — the second marker works without seeing the first marker's marks or comments.
  • Check or sample marking — the second marker reviews a sample of scripts, or the first marker's marks, and confirms or queries them.
  • Moderation — a review across a sample or a whole cohort to confirm that one standard was applied, often ending with an external examiner.

Policies also differ in scope. Full second marking covers every script; sampled second marking covers a defined sample. Whatever the form, the outcome is usually one agreed mark, reconciled between the markers rather than averaged.

A typical handbook mixes them: scripts near a grade boundary double-marked, a sample check-marked, a moderation meeting when agreement slips. The arrangement works, and it is slow. Most of what it reveals about consistency arrives after the marking is over, when the only remedy left is to re-mark.

Can AI be used as a second marker?

Yes, inside a policy, as a documented second read — not as a replacement for the second marker your policy requires. With AI assistance on, NUADU produces a draft read of the same script against the same rubric; a named marker reviews it, overrides where needed and signs, and AI never makes the final grade decision.

The Higher education page puts the fit in one line: authoring, moderation, marking, second marking, appeals — the workflow you already run. AI sits inside it; it does not replace it. How the AI read counts is a policy question, and it belongs to the QA office that owns the policy. One programme might treat it as an extra check on first marking, another as input to the moderation sample, a third as advisory only. Whichever it is, the read leaves something behind: wherever AI assistance is on, there is a draft per criterion to set against the score the marker signed. What the marker sees during that read — a note per criterion and a queue that opens on borderline scripts — is set out in how AI-assisted marking works.

AI is a documented second rater. The marker is still the marker.

Inter-rater reliability: what QA can see

Inter-rater reliability is the extent to which two raters, reading the same work against the same rubric, arrive at the same result. High agreement suggests the rubric is being applied consistently; low agreement points at the rubric, at a rater, or at both. Drift is agreement changing over time — a marker growing harsher as the pile shrinks, or a cohort marked to a slightly different standard from last year's.

In a paper-based policy both stay hidden until the moderation meeting. On NUADU they are visible to QA:

  • Agreement between AI and markers, between markers, and across cohorts.
  • Override rates — how often markers depart from the draft score.
  • Drift indicators, so a shift in standard can be spotted and addressed.
  • Time per response, as a view of the marking, never a label on a learner.

Watch the override rate most closely: it is the number that says whether a second read deserves any trust at all. No agreement figure is offered here, because the figure that matters is the one your QA office sees for your rubric, your markers and your cohort. Where agreement between AI and markers is low, look at the rubric before the markers. If one marker's override rate sits far from everyone else's, that is a conversation to have. The same discipline, applied to items rather than raters, is item analysis.

What documented means

Documented means a file the QA office can request, read and keep on record — not a product page that says so. NUADU's methodology pack covers rubric-application methodology, validity evidence, inter-rater agreement data, bias review, and the marker override and candidate appeals workflows. QA, DPO, procurement and academic-board reviewers can request it.

Beside the pack sit the live records — the trail of inputs, outputs, overrides, approvals and decisions — and what that trail should hold for each AI-assisted decision is the subject of what an AI audit trail records. No AI output reaches a learner without a named educator's approval.

Start from the policy rather than the product. Take the second-marking policy you already run, decide how a documented second read should count in it, and only then look at the agreement data. Marking, rubrics and moderation on the Grade page cover the marking workflow itself, and assessment for higher education shows where second marking sits among everything else a QA office oversees.

→ Talk to us about your assessment workflow (30 min)