Treat an AI-generated question as a draft. Before students see it, verify the learning objective, factual accuracy, answer key, explanation, language and permissions. Keep a human responsible for approval. Fluent wording and a confident answer do not prove that a question is correct or appropriate.
This is an operational checklist for educators, not legal advice or a certification of compliance. Your institute’s policies and applicable requirements still govern student data, content use and assessment. UNESCO’s guidance emphasizes a human-centred approach, privacy and the educational validation of AI tools.
Keep generation and extraction separate
Generating a new question from instructions is different from extracting a question from an existing paper. Generation needs checks on correctness and relevance. Extraction additionally needs comparison against the original: symbols, diagrams, options and answer keys may be altered or omitted. Permission to possess a file does not automatically authorize every reuse.
Review before publication
- Objective: identify the knowledge or reasoning being assessed.
- Scope: check syllabus alignment and prerequisites; do not assume an exam label ensures alignment.
- Correctness: independently solve the item and verify the explanation.
- Options: check that the intended number of answers is correct and distractors are plausible.
- Clarity: remove missing conditions, vague references and accidental clues.
- Notation: inspect equations, units, chemical notation and diagrams in the student view.
- Fairness: remove unnecessary cultural, language or accessibility barriers.
- Permissions and privacy: use approved material and avoid uploading identifiable student records.
- Approval: record who reviewed it, what changed and which version was released.
An example of a small but important correction
Illustrative draft: “A car travels 60 km in 2 hours. What is its speed?” The intended answer is 30 km/h, but “average speed over the journey” is more precise: the information does not show that the car maintained a constant speed. The reviewer should check both the arithmetic and the assumption communicated by the wording.
For a multiple-choice version, inspect every option and the marking configuration. If the explanation says 30 km/h but the saved key points to a different option, a correct question can still produce an incorrect assessment. Preview the published test, not just the editor.
Keep a lightweight approval log
| Field | What to record |
|---|---|
| Question ID and version | A stable identifier and revision date |
| Source or generation input | Approved source reference or non-sensitive prompt |
| Review outcome | Accepted, revised or rejected, with the specific issue |
| Reviewer | The actual educator responsible for checking it |
| Release check | Student preview and answer-key verification |
What to tell students and colleagues
Follow your institution’s disclosure policy and explain how questions are checked. Do not describe unreviewed output as expert-validated or error-free. If an error is found after release, correct the item, review affected scores and communicate the change through the appropriate channel.
ThePrepLab supports AI-assisted question preparation and a separate PDF extraction workflow. Neither removes educator responsibility. Use the lesson-plan template to connect approved questions to an actual teaching objective.