How to Use AI to Create a Quiz (That Actually Tests)
To use AI to create a quiz, give a chatbot your source material plus a five-part prompt: subject, audience, number and type of questions, thinking level, and output format. Paste your notes or upload your document, ask for an answer key with explanations, then run the three checks in this guide before anyone takes it. That last step is the whole job — an unreviewed AI quiz can mark a wrong answer correct, and your students will trust the key more than their own knowledge.
The time at stake is real. A Walton Family Foundation and Gallup survey of 2,232 US public school teachers, fielded in March and April 2025, found that teachers who use AI tools at least weekly estimate saving 5.9 hours per week — the equivalent of about six weeks across a school year. Gallup is careful to note these are teachers’ own estimates, not stopwatch measurements, but the direction is clear: quiz-making is exactly the kind of repetitive task AI handles well. A separate Learnosity survey of 258 US teachers in January 2025 found teachers spend an average of 9.9 hours a week grading — more than a full workday. (Gallup, “Three in 10 Teachers Use AI Weekly, Saving Six Weeks a Year”; Learnosity via Business Wire)
I tested the methods below the way this site tests everything: by building a real quiz from a real source passage and checking every answer against the source by hand. I did not run live signed-in sessions inside ChatGPT, Claude, or Gemini for this guide — the steps, limits, and pricing come from the vendors’ own help pages, which I verified — but the prompt structure and the verification system were worked through on a genuine passage, and the sample questions in this guide are written from it, so you can see exactly what the checks catch.
What Can AI Actually Do When You Ask It to Make a Quiz?
AI quiz generation is two different jobs wearing one name, and picking the wrong one is the most common mistake. A chatbot like ChatGPT, Claude, or Gemini writes questions as text: you get a document you can copy anywhere, but no grading, no delivery, and no analytics. A purpose-built tool like Google Forms with Gemini, NotebookLM, or Wayground builds a quiz you can actually assign: it grades answers, tracks scores, and exports results — but it constrains what you can ask and where the quiz can live.
Start with the job, not the tool. If you need a practice set for yourself, a chatbot is faster. If you need 30 students to take the same quiz and you need their scores, use a tool that grades. Everything in this guide follows from that split.
| Approach | Best for | Answer key | Grading and delivery | Main limit |
|---|---|---|---|---|
| Chatbot (ChatGPT, Claude, Gemini) | Practice sets, study quizzes, quick drafts | Yes, if you ask for one — with explanations | None built in; you copy the text wherever it needs to go | No export, no auto-grading, no score tracking |
| Gemini in Google Forms | Graded class quizzes from a Drive document | Yes — correct answers and points are set automatically | Full Google Forms grading, responses, and export | Paid Workspace or Google AI plan only; desktop only |
| NotebookLM (Gemini Notebook) | Self-study quizzes from your own notes | Yes, with explanations and source citations | Take it on screen; score shown, retakes allowed | No export, no share link, no assigning to a class |
| Dedicated tools (Wayground, ClassPoint, Quillbot) | Classroom delivery, gamified or PowerPoint-based quizzes | Yes, with grading and reports | Built in — shareable links, live play, LMS export | Separate accounts and, for the best features, paid plans |
How Do I Write a Prompt That Produces a Good Quiz?
A weak prompt — “write quiz questions about photosynthesis” — produces generic recall questions that test whether someone read the heading, not whether they understood the chapter. A strong prompt carries five pieces of information, and every one of them changes the output:
- Subject and source. Name the topic and paste the material the questions must come from. Questions grounded in your notes test what you taught; questions from the model’s memory test what it happens to know.
- Audience. “10th-grade biology students” and “new hires in week one” produce completely different questions from the same source. State the level explicitly.
- Number and type. Specify the count and the mix — for example, 6 multiple-choice, 2 true/false, 2 short-answer. Without a mix, you get ten multiple-choice questions by default.
- Thinking level. Ask for recall, application, and analysis questions separately. A quiz that only asks for definitions is a vocabulary list, not a test of understanding.
- Output format. Ask for the answer key separately, with a one-sentence explanation per answer, and ask the model to mark the correct option clearly. This is what makes the verification step possible.
Here is the prompt I use, ready to copy. Replace the bracketed parts with your own material:
“Create a 10-question quiz on [topic] for [audience, e.g. 10th-grade biology students]. Use this source material only: [paste your notes or document text]. Include 6 multiple-choice questions (four options each, one correct answer clearly marked), 2 true/false questions with a brief explanation of why each answer is correct, and 2 short-answer questions that require analysis, not recall. Label each question with its thinking level (recall, application, or analysis). End with a separate answer key, one sentence of explanation per answer. Do not add facts that are not in the source material.”
The last sentence matters most. It is the instruction that stops the model from inventing plausible-sounding facts to fill gaps in your notes — the single most common failure in AI-generated quizzes.
What Does a Good AI Quiz Actually Look Like?
To show the method instead of describing it, I built a sample quiz from a short source passage on photosynthesis — five facts, nothing more. The passage: photosynthesis happens in chloroplasts, mostly in leaf mesophyll cells; chlorophyll absorbs blue and red light and reflects green, which is why leaves look green; the light reactions split water, releasing oxygen and producing ATP and NADPH; the Calvin cycle uses ATP and NADPH to fix carbon dioxide into sugars; and stomata open to take in carbon dioxide and close to conserve water. Here are three questions written from that passage, the way a good AI output should look:
- Recall: Where does photosynthesis mainly take place in a leaf? A) The xylem B) The mesophyll cells C) The root hairs D) The epidermis. Answer: B. The passage places photosynthesis in chloroplasts inside mesophyll cells.
- Application: A plant is kept in a sealed jar with no fresh air for a day. Which process slows down first, and why? Answer: The Calvin cycle slows first, because the plant runs out of carbon dioxide to fix once the stomata cannot take in more — the light reactions can keep splitting water until the ATP and NADPH pile up with nowhere to go.
- Analysis: A student claims the oxygen plants release during photosynthesis comes from carbon dioxide. Is the student right? Explain. Answer: No. The oxygen comes from water split during the light reactions; the carbon dioxide provides the carbon that ends up in sugars.
Notice what the third question does: it tests a misconception directly. The oxygen-from-carbon-dioxide error is one of the most common wrong answers in biology classrooms, and it is also exactly the kind of mistake an AI quiz generator can make in its own answer key. That is why the verification step below is not optional.
How Do I Check an AI Quiz Before Anyone Takes It?
Run three checks on every generated quiz. They take about five minutes for a ten-question quiz, and they catch the failures that damage trust fastest — a wrong answer key erodes confidence in the whole quiz in a single question.
Check 1: Answer-key accuracy. Read every question against your source, not against the AI’s explanation. The explanation can sound convincing while the marked answer is wrong — the model writes the explanation to justify the answer it chose. Verify the answer from the source passage, then read the explanation. In my sample above, the oxygen question is the one to watch: if the key said carbon dioxide, the explanation would probably still sound plausible.
Check 2: Distractor plausibility. Every wrong option should be believable enough that a student who does not know the material could pick it, and wrong enough that a student who does know it will not. The classic failure is a distractor from the wrong domain: “Which pigment makes leaves green?” with options chlorophyll, hemoglobin, melanin, carotene. Hemoglobin and melanin belong to animals, so any student can eliminate them without knowing any biology — the question tests elimination, not knowledge. The fix is to use options from the same family: chlorophyll, carotene, xanthophyll, anthocyanin. All four are real plant pigments; only one makes leaves green.
Check 3: One correct answer. Read each multiple-choice question as a strict test: is there exactly one option that is correct, and are the others unambiguously wrong? Watch for two correct options, for “all of the above” (which lets a student answer without knowing the content), and for options that are correct under a different reading of the question. If a question survives all three checks, it is ready. If it fails one, rewrite the question — do not just fix the key, because a question with a broken structure will keep producing the same failure when you regenerate it.
How Do I Make a Quiz in Google Forms With Gemini?
Google Forms is the fastest route from a document to a graded quiz, with one big condition: the Gemini quiz feature is only available on paid plans. Google announced on July 30, 2026 that the “Help me create” button in Forms can now generate a complete quiz — questions, correct answers, and point values — and that the prompt can reference Google Drive files such as Docs, Slides, or PDFs as sources. The rollout reached Rapid Release domains from July 24, 2026 and Scheduled Release from August 5, 2026. Availability covers Workspace Business Standard and Plus, Enterprise Standard and Plus, Education Plus, and consumer Google AI Pro and Ultra plans. A personal account with no Google AI subscription is not on the availability list. (Google Docs Editors Help, “Generate questions with Gemini in Google Forms”)
The workflow is short. Open a new form on a computer — the feature is desktop-only — and click “Help me create” at the top. Describe the quiz in one sentence, and reference a Drive file if you want the questions drawn from it. Gemini drafts the quiz with answers marked and points assigned, and you preview it before inserting. Then open Settings, turn on “Make this a quiz,” and check every answer key: Google’s own help page tells you to set the answer key for generated questions, and independent testing confirms Gemini occasionally marks a wrong answer as correct. Two more limits to know: the feature works in English only, and only with a single-section form.
If you do not have a paid plan, the same result is one extra step away: generate the questions with the five-part prompt in a chatbot, paste them into a blank Google Form yourself, and turn on “Make this a quiz” to get the same grading and response tracking. The paid feature saves the copy-paste; it does not save the verification.
How Do I Make a Quiz in NotebookLM?
NotebookLM — renamed Gemini Notebook in July 2026, at the same address — is the best option when the quiz is for one person: you, studying your own material. Upload your PDF, notes, or slides as a source, open the Studio panel, and select Quiz. It reads the document and builds a multiple-choice quiz you take right in the app, with difficulty set to easy, medium, or hard, fewer or more questions, and a prompt to focus the topic. Every answer links back to the passage it came from, so a wrong answer is one click away from the explanation — which makes it the easiest tool in this guide to verify against. (NotebookLM)
The catch is what you cannot do next: there is no export, no shareable link, and no way to assign or auto-grade the quiz for a class. It is a self-study tool, not a classroom tool. If you want the same source-grounded method for exam preparation, our guide to using NotebookLM for studying covers the full workflow — one notebook per subject, quizzes after each study session, and the explain-back step for every wrong answer. For the broader exam routine around it, see how to use AI to study for exams.
Which Tool Should I Use for Classroom Quizzes?
When the quiz has to reach other people with scores attached, a dedicated tool earns its place. Wayground — formerly Quizizz — is the classroom standard: gamified live quizzes, a large library of shared quizzes, and reports per student. ClassPoint builds AI quizzes directly into PowerPoint, which suits teachers who already teach from slides. Quillbot’s quiz generator is the lightest option: paste a prompt or a summary, get questions in seconds, copy the text out. None of them replaces the verification step — a dedicated tool grades faster, which makes a wrong answer key worse, not better, because the error reaches every student at once.
The decision rule is simple. One person studying: NotebookLM or a chatbot. One class taking a graded quiz: Google Forms with Gemini if you have a paid plan, otherwise a chatbot draft pasted into Forms. A live classroom game or a PowerPoint lesson: Wayground or ClassPoint. Teachers building a full AI toolkit around quizzes will also want our roundup of the best AI tools for teachers, which covers lesson planning, grading, and feedback alongside assessment.
What Mistakes Ruin AI-Generated Quizzes?
Five mistakes account for most bad AI quizzes, and all five are fixable in the prompt or the review pass. First, generating from memory instead of from a source: without your notes pasted in, the model answers from its training data and invents facts when it is unsure. Second, accepting the first draft: the first generation is a rough cut, and the second pass — “make the distractors harder” or “add two analysis questions” — is where the quiz gets good. Third, testing only recall: a quiz of ten definition questions measures memorization, and students notice. Fourth, ignoring the answer key: most people read the questions and skip the key, which is exactly where the errors live. Fifth, publishing without a trial run: take the quiz yourself, or have one colleague take it, before it reaches a class. A question that confuses the teacher will confuse every student.
One more guardrail for classroom use: tell students when a quiz was AI-generated and that the answer key was human-checked. It sets the right expectation — the quiz is a study aid, not an oracle — and it models the exact behavior you want from them when they use AI themselves.
Frequently Asked Questions
Yes. Upload the PDF to a chatbot like ChatGPT, Claude, or Gemini and use the five-part prompt in this guide, or use NotebookLM, which reads the PDF as a source and builds a quiz with citations back to the passage. Google Forms with Gemini can also reference a PDF in Google Drive, but only on paid Workspace or Google AI plans. Whichever route you take, verify the answer key against the PDF yourself — the AI can misread tables and figures.
It depends on who takes the quiz. For a graded class quiz, Google Forms with Gemini is the fastest complete route if you have a paid plan; otherwise, generate the questions in a chatbot and paste them into Forms. For self-study, NotebookLM is best because every answer links to its source. For a live classroom game, Wayground is the standard. There is no single best tool — there is a best tool for each job.
Yes, with two conditions. First, check every question and the full answer key against your source material before the quiz reaches students — a wrong answer key teaches the wrong fact to the whole class at once. Second, keep student data out of the prompt: paste your teaching material, not student names, grades, or personal information. Most school AI policies allow teacher-created materials; check yours before uploading anything sensitive.
Good enough to draft, not good enough to trust blindly. When the questions are grounded in a source you provide, the questions themselves are usually sound, but the answer key is the weak point — models occasionally mark a wrong answer as correct and write a convincing explanation for it. The three-check system in this guide (answer-key accuracy, distractor plausibility, one correct answer) catches the common failures in about five minutes per ten-question quiz.
Yes, but only inside a tool that grades. A chatbot can write an answer key, but it cannot collect or score responses. Google Forms with “Make this a quiz” turned on grades automatically and exports scores; Wayground and ClassPoint do the same with reports per student. If you only need a practice set for yourself, NotebookLM scores your attempt on screen and lets you retake the questions you missed.
Ten questions is a good default for a practice set or a short class quiz: enough to cover a topic, short enough to review in five minutes. For a formative check during a lesson, five questions is plenty. For a unit test, twenty or more — but generate it in two batches of ten and review each batch separately, because answer-key errors become harder to spot as the quiz gets longer. Quality drops faster than quantity helps, so a tight ten-question quiz beats a sloppy twenty-question one every time.
