An AI-generated quiz can look finished while asking a question the source never answers, offering two defensible choices, or pairing a good question with the wrong answer key. Treat the question, its options and its explanation as separate items to check—not as one approved package.
This workflow is for educators, trainers and anyone preparing a low-stakes practice quiz. The examples below are invented to demonstrate review; they are not outputs from a product test. Important assessments need review by someone who knows the subject and the learners.
Start with the source and the learning goal
Keep the exact passage or lesson version beside the draft quiz. Write a short goal for each question: for example, “identify the standard loan period” or “apply the renewal rule.” A question about a detail never taught may test outside knowledge rather than the intended lesson.
Carnegie Mellon University’s Eberly Center guidance on creating exams emphasizes matching questions to learning objectives, avoiding ambiguity and having another person review the assessment. For multiple-choice items, it recommends a single best answer and plausible alternatives. Those principles provide a useful baseline; the source-to-question checks below make them concrete for an AI draft.
A fictional passage to test against
At Cedar Library, members may borrow up to four books for 14 days. A book may be renewed once for seven additional days if nobody else has reserved it. Reference books must remain in the library. Returns go into the indoor desk slot during opening hours; there is no outdoor return box.
For this exercise, that passage is the entire source. Do not fill gaps with what another library normally does. Keep a review note with four fields: question, source sentence, correct option text, and reason each other option is wrong.
Check whether the source actually answers the question
Consider: “What fine does Cedar Library charge for a late book?” Even a confident answer such as “$1 per day” is unsupported here. The passage does not describe fines. A better explanation cannot fix the missing evidence.
Either remove the item or obtain an authorized source that includes the policy and make that source part of the lesson. Do not invent a fine simply to keep the quiz at a requested length. Apply the same check to names, dates, exceptions and definitions introduced by the answer explanation.
Try to prove every option right
A draft asks: “Which books cannot be renewed?” Its choices include “reference books” and “books reserved by another member,” with the second marked correct.
The problem is the scope: reference books do not leave the library, so a learner could reasonably select them too. The question does not make clear that it concerns a borrowed book. Rewrite the stem:
A member has borrowed a book and has not renewed it before. Which condition prevents renewal under Cedar Library’s policy?
- A. Another member has reserved the book.
- B. The member originally borrowed four books.
- C. The member wants seven additional days.
- D. The book’s original loan period was 14 days.
Answer: A. The reservation condition blocks renewal. B describes an allowed borrowing quantity; C matches the permitted extension; D matches the standard loan period. The added wording also rules out a previous renewal without turning that exception into an accidental second answer.
Audit the answer key independently
Now imagine the export labels D as correct but keeps the explanation about a reservation. The prose may sound sensible while the scoring is wrong. Compare the key with the text of the correct option, not just its letter.
After any option reorder, check the displayed letter, stored correct choice and feedback again. If the quiz platform shuffles answers, preview that behavior too. A spreadsheet with “A” in the key column can become stale when someone rearranges the choices elsewhere.
Make wrong answers useful, not misleading
For “How long is the standard book loan?”, seven days is a useful wrong choice because it confuses the extension with the original period. “Until a spaceship arrives” adds little: a learner can reject it without understanding the policy.
A distractor should be incorrect for a clear reason, but plausible enough to reveal a relevant misunderstanding. Check that none becomes correct under an unstated assumption. Avoid making the answer obvious through an unusually detailed choice or a phrase copied only into the correct option.
Also read the question without looking at the choices. If you cannot tell what it asks, repair the question rather than hoping the options explain it. Use the same vocabulary learners encountered in the lesson unless interpreting a new term is itself the goal.
Review the learner’s version before sharing
- Hide the key. Ask a colleague to answer from the provided material and flag any item they can defend in more than one way.
- Inspect feedback. Explanations should point to the relevant rule, not introduce unsupported extra policy.
- Preview the actual delivery. Check that options, punctuation, images and any required source passage remain visible after export or import.
- Keep a correction record. Save the approved source version and answer text alongside the quiz. Recheck affected items when the lesson changes.
Tools such as Quizgecko can help prepare quiz drafts, but generation does not replace these checks. If your task is collecting opinions rather than assessing knowledge, use a different review approach: our guide to checking AI-written survey questions focuses on wording and respondent interpretation rather than a correct-answer key.
A shorter quiz with defensible questions is more useful than a longer one padded with unsupported items. Approve the evidence, wording and scoring together before asking learners to rely on the result.
