AI and studying

How to use AI for multiple-choice exams

AI can write you a hundred practice questions before your coffee cools. Whether they help depends on how you use them, and on whether you check them against what you were actually taught.

By the Ace team · November 16, 2026 · 6 min read

How to use AI for multiple-choice exams — Ace blog

Short answer: use AI to write practice multiple-choice questions from your own lecture material, answer them from memory before you look at anything, and check every answer key against your notes. Studies so far find AI-written questions roughly as good as human-written ones overall, but some have two right answers or a wrong option that gives the game away, so treat the AI as a question writer you supervise, not an answer source.

Why multiple-choice practice is worth doing

Multiple-choice questions have a reputation as the lazy option: you only have to recognise the answer, not produce it. The research is kinder than that, with one condition. In two experiments, Little, Bjork, Bjork and Angello (2012) found that well-built multiple-choice tests triggered real retrieval, and that practising with them even helped people answer related questions about the wrong options, something cued-recall practice did not do. The condition is that the wrong options have to be plausible. If one choice is obviously silly, you never have to think about why the right answer is right.

That is exactly where AI can help. Writing four believable options for every question is slow and hard, which is why most students never make their own multiple-choice practice. An AI tool can draft them in seconds. Your job is to make sure the drafts are good, and to use them in a way that makes you remember rather than read.

If your exam is mostly multiple choice, pair this with our guide on how to study for multiple-choice exams; this post is about the AI part.

Are AI-written multiple-choice questions any good?

Better than you might expect, but not good enough to use unchecked. Three studies give a fair picture:

  • Medical exams. Cheung and colleagues (2023, PLOS ONE) had ChatGPT and two university examiners each write 50 graduate-level medical questions, then asked blinded assessors to rate them. Total quality scores were not significantly different. The one area where the AI did worse was relevance to the syllabus. ChatGPT took about 20 minutes for its 50 questions; the humans took over three and a half hours between them.
  • Programming courses. Doughty and colleagues (2023) compared 651 GPT-4 questions with 449 human-written ones across six Python courses and found GPT-4 could produce questions with clear language, a single correct choice and good distractors, well matched to the course's learning objectives. A University of Warwick summary of the project still flags the two classic failures: some AI questions had more than one correct answer, and some distractors revealed the right one.
  • Actual learning. An and colleagues (2025) gave about 60 data science students a week with LLM-generated multiple-choice retrieval practice and a week without. Average accuracy was 89% after the practice week and 73% after the week without it. The authors are clear that question quality varies and that each question needs checking before students see it.

Put together: AI is a fast, decent question writer whose weak spot is staying on your syllabus and getting every key right. Both problems shrink when you give it your own material and check its work.

How to use AI for a multiple-choice exam, step by step

  1. Give it the source, not the topic. Upload or paste the lecture slides, notes or chapter you will be examined on. "Questions on the cell cycle" gets you the internet's version; "questions from these slides" gets you your lecturer's. This is the fix for the relevance gap Cheung's assessors spotted.
  2. Ask for one question at a time. A list of twenty questions with the answers underneath turns practice into reading. One question, then wait, keeps you retrieving.
  3. Commit before you see the options. For a harder workout, ask for the question stem first, say your answer out loud, then ask for the four options. If your answer is among them, you knew it; if not, you were relying on recognition.
  4. Explain the wrong options. After each answer, say why each distractor is wrong. This is the part Little and colleagues found carries over to related questions.
  5. Keep a misses list. Copy every question you got wrong into a document and redo it in two or three days. Spaced, repeated attempts beat one long session.
  6. Finish with a timed set. In the last week, ask for a mixed set across topics and answer it against the clock, so you practise switching topics as the real paper will make you.

Prompts that produce useful questions

  • "Using only the material I've uploaded, write one multiple-choice question with four options. Make every wrong option something a student who half-learned this section would pick. Don't show the answer until I reply."
  • "After I answer, tell me if I'm right, quote the sentence from my material that proves it, and explain in one line why each other option is wrong."
  • "Write a question that tests applying this idea to a new example, not recalling a definition."
  • "Here are the questions I got wrong last time. Write a new question on each idea, worded differently."
  • "Check this question: is exactly one option correct according to my notes? Does any wrong option give away the answer?"

The quoting step matters most. If the tool cannot point to a line in your material, treat the question as suspect.

How to check an AI question before you trust it

Run each question through four quick checks, the same faults the studies above found:

  • One right answer. Could a second option also be defended from your notes? If so, the question teaches confusion. Fix it or drop it.
  • No giveaways. Is the correct option longer, more detailed or the only one using a word from the stem? Real exams avoid these clues; your practice should too.
  • On your syllabus. Does the question test something your course actually covered, at the depth it covered it?
  • Key matches the source. Find the sentence in your own material that supports the answer. If you cannot, look it up before you learn it. Our guide on whether you can trust AI study answers explains why chatbots can be confidently wrong.

This checking is also why some tools build it in. Ace (acenow.app) is a study app that turns your own lecture PDFs, slides or YouTube lectures into multiple-choice questions checked against the source. Every question is written from your material only, questions that fail the check are dropped, and under each answer Ace shows the sentence from your file that backs it. You can also duel a friend on the same questions. See how it works.

Frequently asked questions

Can ChatGPT make a multiple-choice practice test from my notes?

Yes. Upload or paste your notes and ask for questions drawn only from them, one at a time. Ask it to quote the line that supports each answer, and check that line yourself.

Are AI multiple-choice questions as good as a teacher's?

In the studies so far, overall quality was similar, but AI questions were slightly less relevant to the syllabus in one medical study, and some had more than one correct answer or a distractor that gave the answer away. Use them, but check them.

Is multiple-choice practice enough on its own?

It works best alongside some recall practice, such as answering before you see the options or explaining why each wrong option is wrong. If your exam also has written questions, practise those in their own format.

Should I use AI during a multiple-choice exam?

No. Use it to prepare, not during an assessment, unless your course explicitly allows it. Check your institution's rules on AI before using it for any graded work.

Sources

  1. Little, J. L., Bjork, E. L., Bjork, R. A., & Angello, G. (2012). Multiple-choice tests exonerated, at least of some charges: Fostering test-induced learning and avoiding test-induced forgetting. Psychological Science. Summary by the Association for Psychological Science
  2. Cheung, B. H. H., Lau, G. K. K., Wong, G. T. C., Lee, E. Y. P., Kulkarni, D., Seow, C. S., Wong, R., & Co, M. T.-H. (2023). ChatGPT versus human in generating medical graduate exam multiple choice questions: A multinational prospective study. PLOS ONE, 18(8), e0290691. https://pmc.ncbi.nlm.nih.gov/articles/PMC10464959
  3. Doughty, J., Wan, Z., Bompelli, A., et al. (2023). A comparative study of AI-generated (GPT-4) and human-crafted MCQs in programming education. arXiv:2312.03173. https://arxiv.org/abs/2312.03173
  4. University of Warwick. A comparative study of AI-generated (GPT-4) and human-crafted MCQs in programming education (project page). https://warwick.ac.uk/fac/cross_fac/eduport/edufund/projects/yang/projects/a-comparative-study-of-ai-generated-gpt-4-and-human-crafted-mcqs-in-programming-education/
  5. An, Y., Liu, J., Acharya, N., & Hashmi, R. (2025). Enhancing student learning with LLM-generated retrieval practice questions: An empirical study in data science courses. arXiv:2507.05629. https://arxiv.org/abs/2507.05629

Turn your notes into practice questions

Drop a PDF, slides or a YouTube lecture into Ace and get practice questions checked against your own material — then play them solo or against a friend.

Try Ace free