Creating Quizzes and Exam Questions with AI

AI can speed up assessment design: generating question pools, producing variants, drafting rubrics, and helping you review items for clarity. But it only becomes good pedagogy when you use it to support a deliberate assessment workflow: clear learning outcomes, appropriate cognitive level, and careful review for accuracy and fairness.¹

Well-designed quizzes also serve learning (not just grading). Frequent retrieval practice (“practice tests”) reliably improves long-term retention, especially when paired with feedback.^{2 3 4}

A Workflow That Works (and Scales)

Prompt Templates (Copy/Paste)

Use these as “starter scaffolds”. Replace bracketed text.

1) Blueprint first

You are an assessment designer for a university course.
 
<course_context>
Discipline: [e.g., Intro economics]
Learners: [e.g., 1st-year undergraduates]
Assessment: [quiz/exam], duration: [e.g., 45 minutes]
Allowed resources: [closed book / one page notes / open book]
</course_context>
 
<learning_outcomes>
- [Outcome 1]
- [Outcome 2]
- [Outcome 3]
</learning_outcomes>
 
<constraints>
- Coverage: include all outcomes, emphasize [topic X]
- Cognitive level targets: [e.g., 30% remember/understand, 50% apply, 20% analyze]
- Item mix: [e.g., 8 MCQ, 3 short answer, 1 problem]
- No trick questions; avoid cultural references and idioms
</constraints>
 
Task:
1) Create an assessment blueprint table with outcomes × item type × Bloom level × number of items × points.
2) Briefly justify the blueprint in 5–7 bullet points.

2) Generate MCQs with rationales and misconception-based distractors

Research-based tip: three-option MCQs are often sufficient and can reduce time spent inventing weak distractors.⁶

Generate 12 multiple-choice questions aligned to this blueprint:
[paste blueprint table]
 
Rules:
- Prefer 3 options per item unless there is a strong reason for 4.
- Provide: stem, options (A–C), correct answer, and a 1–2 sentence rationale.
- For each distractor, name the likely misconception it represents.
- Avoid "all of the above" and "none of the above" unless necessary.
- Keep reading load minimal; test the concept, not English fluency.
 
Return as a table.

3) Audit and revise (AI as editor)

Review the following questions for quality.
 
Checks:
- Alignment to the intended learning outcome and Bloom level
- Clarity and single-interpretation stems
- Construct-irrelevant difficulty (unnecessary complexity, tricky wording)
- Bias and accessibility (idioms, culturally specific knowledge, screen-reader issues)
- For MCQs: plausibility and distinctiveness of distractors
 
For each item:
1) Give a brief issue list (if any)
2) Propose a revised version
3) Explain what changed and why
 
Questions:
[paste questions]

Question-Type Playbook (What to Ask AI For)

Multiple-choice questions (MCQ)

Use MCQs when you want breadth of coverage and efficient grading. Strong MCQs can also assess application and analysis, not only recall.^{6 7}

Ask AI to draft stems from scenarios or cases (application), generate distractors based on known misconceptions, write brief rationales for correct and incorrect options (useful for feedback and review), and spot unintended cues (grammatical agreement, option length patterns, absolute words).

Short-answer and problem-solving

Use short answers when you want students to produce a response (not recognize it). They also work well for partial credit.

Ask AI to create a scoring guide (key steps, points per step), propose common wrong turns and partial-credit rules, and generate 3–5 equivalent variants (numbers or context changed, same method).

Essay questions

Essay prompts are powerful when you want synthesis, evaluation, and argumentation, but they require clear criteria and calibration.

Ask AI to draft prompts that force tradeoffs (not “tell me everything”), create a rubric aligned to outcomes and cognitive demand, and generate anchor responses plus a grader calibration checklist.

Quality Checklist (Use Before You Publish an Exam)

Alignment: Every item should map to a learning outcome and the intended Bloom level.⁵

Coverage: The full set should match your blueprint, avoiding over-testing one lecture.

Clarity: Students should be able to interpret each question in one reasonable way.⁹

Fairness & accessibility: Remove construct-irrelevant barriers and apply UDL principles (multiple ways to demonstrate learning, reduced irrelevant reading load, clear structure).¹⁰

Security: Remove details that reveal future assessments, and use variants or rotate pools when appropriate.

Scoring: Ensure rubrics and answer keys are explicit enough that another grader could apply them consistently.¹

Academic Integrity, Privacy, and Transparency

Follow your institution’s policy for AI tools, and be explicit with students about what is allowed for studying versus graded work.

Avoid pasting confidential exam banks, student work, or identifiable student data into third-party tools unless approved.

If you use AI to generate assessment items, keep a “paper trail” from outcomes → blueprint → items → revisions. This supports quality review and defensibility.¹

For inclusive practice, prioritize proactive design (clear expectations, accessible formats) rather than relying only on accommodations.¹⁰

For broader guidance on responsible use of generative AI in education (including privacy and governance considerations), see UNESCO’s guidance.¹¹

References & Footnotes