Question-bank integrity lab

# AI quiz duplicate-question detection test

A duplicate classification matrix and candidate-pair review protocol for separating exact copies, formatting variants, templated near-duplicates, semantic equivalents, related practice, and intentional parallel forms.

Published18 October 2026Reading time12 minEvidenceDuplicate classification matrix and candidate-pair review protocolAuthorThe AI Quiz Lab

Detect duplicates before adding AI-generated questions to a bank. Compare stable IDs, normalized stems, answer sets, evidence targets, learning objectives, and semantic intent. Separate exact copies, formatting variants, templated near-duplicates, semantic equivalents, and legitimately related practice. Similarity is a retrieval signal, not an automatic deletion rule.

## Key findings, verified 18 October 2026

- Moodle's current question-bank documentation separates reusable shared banks from quiz-specific banks and exposes question identity, versions, status, usage, and review information.
- Canvas currently lets authors duplicate a New Quizzes question, including its stem, answers, points, options, and aligned outcomes; deliberate copying can therefore create valid starting points as well as accidental repetition.
- The QDup research paper treats near-duplicate detection and related-question retrieval as distinct tasks and uses a hybrid statistical and neural pipeline to surface candidate pairs.
- A similarity score cannot determine whether two items are redundant, intentionally parallel, or merely related; the evidence target and version purpose still need human review.
- Deduplication does not prove source accuracy, answer correctness, coverage, fairness, accessibility, reliability, or assessment validity.

## What counts as a duplicate quiz question?

A duplicate is an item that measures the same target with materially the same evidence and response path, even if superficial wording changes. Exact copies are the easiest case. Harder cases include reordered options, renamed variables, paraphrased stems, shared answer rationales, and scenario shells that preserve the same calculation. Related items may share a topic while asking for different evidence and should not be removed automatically.

## Duplicate classification matrix

ClassSignalExampleDefault actionExact duplicateSame stable content or checksumIdentical stem, options, key, and explanationKeep one canonical record after reference checksFormatting duplicateOnly punctuation, spacing, markup, or irrelevant option order differsSame item after two importsMerge after a field-by-field comparisonTemplated near-duplicateSame reasoning with substituted names or numbersIdentical calculation shellReview exposure and purposeSemantic duplicateDifferent wording asks for the same evidence“Define X” and “What is X?”Keep the stronger traceable itemKey collisionDifferent stems reuse one unsupported answer or rationaleSeveral items repeat one explanationRe-audit evidence and keysRelated itemSame topic, different objective or evidenceDefine a rule versus apply itPreserve when coverage is intentionalParallel formEquivalent construct with controlled variation and declared purposeVersion A and B itemsPreserve with form metadata

## Reproducible duplicate-detection test

- Export the complete bank. Include stable IDs, source IDs, objectives, response types, options, keys, explanations, versions, and quiz references.
- Normalize superficial features only. Standardize case, repeated whitespace, punctuation, markup, and option order only where order is not meaningful.
- Find exact matches. Compare normalized stems, answer sets, keys, and explanations; do not rely on the stem alone.
- Retrieve candidates. Use token overlap and semantic similarity to create a review queue, and store the method and threshold as retrieval settings rather than verdicts.
- Compare evidence targets. Review source evidence, objective, cognitive action, response format, key, scoring, explanation, and intended form.
- Classify each pair. Label it exact, formatting, templated, semantic, related, parallel, or distinct and record the reason.
- Select a canonical record carefully. Recheck the key, source, accessibility, usage, and version history before choosing one item.
- Redirect references. Update every quiz, pool, branch, score rule, and analytics reference before archiving a redundant record.
- Rebuild coverage. Confirm that removal did not create objective gaps, shrink a random pool below its requirement, or unbalance a form.
- Run regression controls. Keep known duplicate, known related, and known parallel pairs and rerun them whenever generation, embedding, normalization, or import logic changes.

## Worked example: similar wording, different evidence

Two questions ask for the “best next action” after a security alert. One uses a verified incident-response policy and tests escalation; the other uses a customer-support playbook and tests communication. Text similarity is high, but the evidence, objective, and correct action differ. They are related, not duplicates. By contrast, a paraphrase that changes only the employee name while preserving the same policy clause and response path is a templated near-duplicate.

## Candidate-pair review specification

For every candidate pair, return both stable item IDs, similarity signals, source evidence IDs, objective IDs, cognitive actions, response formats, keys, explanations, version purposes, proposed class, reviewer rationale, canonical decision, and affected quiz references. Never archive an item solely because an embedding or text threshold was exceeded.

## Publication blockers

- Items lack stable IDs, source IDs, or a declared version purpose.
- Normalization removes a meaningful number, negation, unit, or order cue.
- A similarity score is treated as a duplicate verdict without human comparison.
- Parallel-form, accommodation, or practice items are merged accidentally.
- Removing a record breaks objective coverage, pool size, randomization balance, score rules, or analytics continuity.
- Quiz references still point to an archived item.
- The regression set contains no known duplicate, known-related, and known-parallel controls.

## What this test does not prove

The test retrieves and classifies overlapping items. It does not prove that the retained item is accurate, accessible, fair, reliable, or valid. It also cannot establish that parallel forms are psychometrically equivalent; that requires designed evidence and appropriate analysis. QDup is evidence that hybrid retrieval can surface candidates, not a universal threshold or an automatic publication decision.

## Sources, verification date, and corrections

This guide was verified on 18 October 2026 against Moodle's current [question-bank documentation](https://docs.moodle.org/en/Question_banks), Canvas's current [question-copy documentation](https://community.instructure.com/en/kb/articles/661065-how-do-i-duplicate-an-individual-question-in-new-quizzes), and the [QDup research paper](https://arxiv.org/abs/2301.05150) on unsupervised near-duplicate and related-question retrieval. Those sources support question reuse, intentional copying, and candidate retrieval as a separate problem. The classification matrix, ten-step test, worked example, review specification, and blockers are original AI Quiz Lab evidence assets. Report changed guidance or reproducibility problems through the [corrections process](https://bestaiquizgenerator.com/corrections).

## Next checks

After deduplication, rerun the [learning-objective coverage matrix](https://bestaiquizgenerator.com/blog/ai-quiz-learning-objective-matrix), [answer-key checklist](https://bestaiquizgenerator.com/blog/ai-quiz-answer-key-checklist), [randomization test](https://bestaiquizgenerator.com/blog/ai-quiz-randomization-test), and [item-analysis plan](https://bestaiquizgenerator.com/blog/ai-quiz-item-analysis).

---

Canonical: https://bestaiquizgenerator.com/blog/ai-quiz-duplicate-question-test