Create · AI-generated exam questions: drafts, not items

AI-generated exam questions are drafts, not items: on NUADU a draft cannot reach a learner until a named educator has reviewed and approved it. This post walks that workflow end to end for heads of assessment and programme directors: the inputs you already hold, the three ways an item enters the bank, the review pass and the failure modes it exists to catch, and the record every approved item carries. Written by hand, imported or drafted with AI assistance on, every item reaches the same gate.

Your materials in, candidate items out

The workflow does not start with a prompt. It starts with what a programme already has: a rubric, a framework, and the materials it teaches from — readings, course notes, training documents. Those are the inputs the exam authoring workflow on the Create pillar is built around. With AI assistance switched on, NUADU's Content Scanner drafts candidate items from your materials and proposes a question type, a level and a draft rubric for each. Every candidate item is held for educator review.

If you were expecting a prompt library, the Content Scanner works from your materials instead. The output is a set of candidates for a named educator to judge against your rubric and framework. The Create page puts the division of labour in two sentences: "Authoring takes judgement. NUADU shortens the drafting, not the judgement."

Two boundaries keep the drafting honest. AI is opt-in per institution, per programme and per assessment, so a department that wants manual authoring only never meets the drafting assistant. And AI never publishes. On the AI in assessment page, the Content Scanner's human/never pair reads "Approve, edit or reject every drafted item" against "Publish an item". To see what it proposes before deciding whether to switch it on, try the Content Scanner's live demo.

A drafting assistant. Never a publisher.

Three ways in, one gate out

There are three paths into a NUADU bank, and two of them involve no AI at all. The Create page's headline names them: "Author by hand. Import what you have. Use AI as a drafting assistant — your choice."

• Author in the editor, with rich media, across 24+ question types.
• Import from an existing bank, with metadata, framework mapping and rubric structure — the route our post on item bank migration covers.
• Switch on AI drafting and review what the Content Scanner proposes.

All three end at educator review and approval. The educator edits, reorders or rejects any item, adjusts rubric criteria and weights, confirms the framework alignment manually and approves. The author of record is the approving educator, not the drafting tool.

One thing should happen before any of this. On NUADU the rubric is anchored to the item from the start and travels with it into marking and feedback, so the hour spent writing the rubric before the items is the best-spent hour in the process. A rubric the AI drafts is a starting point; the educator finalises it.

How do you review AI-generated exam questions?

Review them the way you would review a colleague's draft: item by item, against the rubric criterion each one claims to evidence, before any learner sees it. A draft can be fluent and still wrong for its purpose. Three failure modes are worth checking on every item.

• Recall where the rubric wants analysis. The stem asks for a definition; the criterion asks for a judgement. Rewrite the stem, or move the item to a criterion it actually evidences.
• A distractor that gives the answer away. An option that is longer, more precise, or the only one that reuses the stem's wording tells the candidate where to click. Rewrite the distractors, or reject the item.
• An alignment proposal that overreaches. With AI on, the Standards mapper proposes an alignment to Bloom's, CEFR, EQF or your own framework, and a proposal becomes the alignment only once a human confirms it. If the item evidences less than the tag claims, change the tag; our post on assessment alignment explains why the tag is a judgement.

None of these is caught by trusting the draft. They are caught by a named educator reading the item against the rubric, which is the reason the gate exists. Where a bank runs in more than one language, translations go through the same gate: reviewed by named educators, not auto-published.

What the bank records

Once approved, an item carries a named approver and an audit record. The bank is versioned — items, named approvers, change history — and ready for QA and external review. For framework alignment, the audit record shows the proposal, the change (if any) and the named approver, so an external examiner sees not only what an item is aligned to but who agreed and what they changed.

That record is the point of the workflow. An appeal panel, a quality auditor and an accreditation reviewer will each ask of an item what they ask of a grade: how was it decided, and by whom? A bank in which every item was approved by a named educator has answered before anyone asks. Every AI output along the way is logged and reviewable on request.

From the bank, the item flows into Assign, Conduct, Grade, Report and Certify with the same rubric and framework alignment it was approved with. Nothing enters the bank unreviewed, and every item in it already carries its approver and its record.

Frequently asked questions

Can AI generate exam questions?

As drafts, yes. With AI assistance on, the Content Scanner drafts candidate items from your materials and proposes a question type, a level and a draft rubric for each. A named educator approves, edits or rejects every one before it can reach a learner.

What is the best AI question generator?

For an institution, the more useful question is what happens to the output. On NUADU the answer does not depend on the drafting step: every candidate is held for a named educator, read against your rubric and framework, and approved with an audit record. Manual authoring and bank import need no AI at all.

How do you evaluate test questions?

Before use, a named educator reads each item against the criterion it claims to evidence and checks the stem, the distractors and the alignment tag. After use, item statistics such as difficulty and discrimination show how it performed — the subject of our post on item analysis in assessment.

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