AI-assisted marking: the marker decides, rubric for both
AI-assisted marking, explained: a draft score with a note per rubric criterion, borderline scripts first, and a named marker who overrides and signs.
AI-assisted marking is marking in which software drafts a score and an explanation against your rubric, and a named marker reviews the draft, overrides where needed and signs the result. On NUADU, switching AI assistance on gives each essay or spoken response a draft score and a rubric-anchored explanation, and AI never makes the final grade decision. For a head of assessment weighing the switch, one idea carries the rest: the explanation matters more than the score. A number with no reason per criterion cannot be checked, overridden with a reason or defended on appeal.
Can AI grade essays with a rubric?
It can draft a grade against one. It should not decide one. With AI assistance on, NUADU applies your rubric to each written response and returns a draft score with a note against every criterion, flags borderline cases, and shows where it is less confident. The marker reads the draft, changes whatever the response does not support, and signs.
The rubric is the important word in that question. The draft is anchored to the criteria your institution wrote, not to a generic model's opinion, and the same rubric runs through authoring, marking and feedback. That puts the weight on the rubric itself: a vague descriptor yields a vague draft, which is why writing the rubric comes before writing the items. The same question asked of a general-purpose chat tool has a different answer, set out in whether a general chat assistant can grade essays.
The limits are written down as well. NUADU's AI Transparency Notice names them without softening: a model can mis-score a response, miss an error or produce feedback that does not fit, and nothing it outputs is definitive until a person has reviewed it. Read that as a design requirement rather than a disclaimer — it is the reason the workflow ends with a person.
Why the explanation matters more than the score
A bare draft score leaves the marker two moves: accept it on trust, or mark the script again from the beginning. Neither is assistance. A draft with a note per criterion — where the response met the descriptor, where it fell short, and where the draft is less confident — adds a third move. The marker checks where the draft is unsure or where criteria pull in different directions, and moves faster where they agree. The review becomes targeted instead of total.
Defensibility runs the same way from the other end. An override needs a reason, and a reason needs a criterion to attach to. On NUADU every override is logged, and the entry has a plain shape: the criterion, the score before and after, the reason, and on the signed result the marker's name and a timestamp. A QA office can read that. A total that changed with no criterion behind it tells a reviewer nothing.
The candidate sees the same structure: feedback tied to the rubric, the marker's notes, and a result broken down criterion by criterion. The institution's appeals process applies. A dispute then has something to test: which criterion moved, who moved it, and why.
Marking is a bottleneck. Removing it cannot mean removing the marker.
Which scripts does the marker see first?
The ones where a marker's judgement changes the outcome. NUADU's Marking assistant starts when an exam is submitted. It moves borderline cases, and responses whose scores vary widely from one rubric criterion to the next, to the top of the marker's queue, and it marks where its draft is confident and where it is not. Its output goes to the marker, and the educator signs off.
Queue order sounds administrative. It is a quality decision. In a typical marking run many responses are clear — clearly strong, clearly weak, clearly mid-band. A smaller number sit on a grade boundary, score very differently on different criteria, or answer in a way the rubric did not anticipate. A queue in arrival order spends the marker's freshest attention at random and leaves the hardest scripts for the last tired hour. A borderline-first queue puts that attention where it decides something.
What an AI marking assistant must not do is just as specific. It never releases a mark on a flagged case without the marker. Confident drafts are not released either: confidence is information for a person about where to look harder, and the marker reviews and signs off every grade. No time-saved figure is offered here, because the case for ordering a queue rests on where judgement goes, not on a stopwatch.
Can teachers use AI when marking essays?
Yes — as an assistant that drafts, not as the one who decides. The pressure behind the question is familiar. Feedback arrives too late to inform learning. Marker consistency drifts across the cohort. Higher-order feedback is the first casualty when time is short.
The tempting answer is AI marking that replaces the marker. It puts a high-stakes academic decision behind a model that nobody on staff can fully explain. That is not a fix — it is a transfer of risk. The Grade page names the alternative: AI as a documented second rater, with the marker still signing off, gets you the operational improvement without the transfer of risk. The signed loop keeps the order fixed. Marking starts when the exam is submitted, the educator reviews when draft marks are ready, and feedback is released only once marks are signed off.
Two questions follow, and each has its own answer. How a documented second read fits the policy your QA office already runs, and what agreement data it makes visible, is the subject of second marking with an AI read. What the learner receives after sign-off is rubric-based feedback with the marker's name on it.
What to consider before switching on AI-assisted marking
AI-assisted marking is a configuration decision, not a platform default. Six points for the conversation with your QA officer:
- Scope. AI is opt-in, module by module: on for one assessment, a programme or the whole institution, or not at all — and the platform keeps working with it off.
- Your rubric. A draft is only as specific as the criteria it is anchored to.
- Languages. Which languages AI-assisted marking covers is confirmed per deployment, not assumed.
- Spoken responses. They are transcribed with playback and scored per criterion; the marker reviews at audio level, and the AI never finalises a speaking grade.
- The record. Each AI output, override and sign-off lands in the audit trail, where QA review and candidate appeals can find it.
- The regulation. The EU AI Act treats AI used to evaluate learning outcomes as high-risk, with obligations around human oversight, logging and documentation. NUADU's "EU AI Act-ready" describes those controls built into the workflow — not a certification claim.
Bring a sample of marked work and walk through the override workflow with your QA officer. The Grade page on marking, rubrics and moderation sets out the workflow end to end, there is a live demo of AI-assisted marking for written responses, and NUADU's AI capabilities lists what the educator decides and what the AI never does, module by module.