Review AI-generated quiz questions before you host

Choose the audience, the learning target, and the material the quiz should cover. A request such as “five questions about safety” leaves important context unspecified. A stronger brief identifies the approved procedure, the role of the learner, and the particular decision they need to practise. Do not ask the draft to stand in for a policy or a subject expert.

UNESCO’s guidance describes a human-centred approach to generative AI in education, including institutional validation and privacy. Apply that principle by letting a person select the objective and approve the output. This is an editorial recommendation, not a claim that AI improves learning or guarantees a faster review.^[2]

Check permission before sharing material with the service. Remove learner records, customer details, passwords, and confidential examples unless your organisation explicitly permits that use. Use neutral fictional details in practice scenarios. A useful question rarely needs someone’s private information.

NIST identifies data privacy risks; UNESCO also highlights user-data protection. Check your institution’s rules before use. These sources do not establish the terms of your account or provider. If material cannot be shared appropriately, write the question yourself or use an approved alternative.^[1][2]

Hide or ignore the marked answer and solve the item yourself. Recalculate quantities and check units, assumptions, and rounding. Then compare your result with every option. A second prompt asking the same system whether it is right is not independent verification; the system may repeat the same mistake or produce a convincing explanation for it.

NIST’s generative AI profile describes confabulation: confidently presented false or erroneous content, including misleading logic or citations. That is why the review should test the underlying reasoning rather than the tone of the explanation. For consequential or specialist material, involve a qualified reviewer instead of relying on a general-purpose quiz workflow.^[1]

If a question depends on a policy, scientific statement, historical fact, or current rule, locate an authoritative source and check what it actually says. A source should support the precise claim, including its date and context. A real publication can still be irrelevant to the proposed answer, and a plausible-looking title or URL may not identify a real source at all.

1. Open the original source rather than accepting a generated summary as the evidence.
2. Check who published it, when it applies, and whether it covers your audience or jurisdiction.
3. Find the passage or calculation that supports the answer.
4. Record a brief source note for the person who will facilitate the quiz.
5. Remove or rewrite the item if the answer cannot be established with confidence.

NIST explicitly includes fabricated citations among the ways generated content can mislead. Keep source notes with your authoring materials; a generated bibliography does not become reliable simply because it is formatted like an academic reference.^[1]

A correct answer does not rescue a confusing question. Check that the stem gives the full problem, only one option is defensible, and the incorrect choices are plausible without being ambiguous. Waterloo’s multiple-choice guidance recommends clear stems, familiar language, and meaningful distractors. These checks apply equally to human-written and AI-assisted drafts.^[3]

• Remove unnecessary negative wording, double negatives, and clues created by answer length.
• Make dates, units, and conditions explicit where they affect the answer.
• Remove stereotypes or personal details that do not serve the learning target.
• Ask whether selecting an option demonstrates the skill you intend to practise.
• Add a discussion or practical task when recognition alone is not enough.

Review the questions in the quiz editor after accepting a draft. OpenQuiz uses 2–6 options with exactly one correct answer and a 5–300 second timer per question. Confirm the correct marker, question order, and time for reading and thinking. A structurally valid quiz can still contain a wrong answer, an unfair assumption, or material that is unsuitable for the audience.

Finish corrections before creating the lobby. An existing OpenQuiz room keeps the question version with which it was created; editing the source quiz does not replace the questions in that room. If you discover an error after creating a room, correct the quiz and create a new room before inviting the group to play. This prevents an old answer key from remaining in the session you intend to host.

Use a final gate rather than assuming the draft is ready. Hold back any item that fails a check below. You do not need to use every generated question; rewriting a weak item is often clearer than adding caveats.

• I can explain why the marked answer is correct without relying on the generated key.
• I checked the necessary sources, permissions, and context.
• I can explain why each other option is incorrect under the stated conditions.
• The wording, difficulty, and timing are appropriate for this group.
• The saved quiz and the room I will host contain the reviewed version.

Involve the host: they need to explain the answer and respond to reasonable challenges. AI can contribute wording; responsibility remains with the author and facilitator.