# How to make a quiz from a PDF, and check every question

**How do I make a quiz from a PDF?** To make a quiz from a PDF, get the text out of the file, keep the one chapter or section that matters, pick question shapes that fit the content, and have the questions written or write them yourself. Then check every question against the PDF. The answer must be on a page you can point to, the key right, and only one option correct.

Published 2026-10-09 by EdProcto. https://edprocto.com/guides/how-to-make-a-quiz-from-a-pdf

You can make a quiz from a PDF in a few minutes. Finding out whether the quiz is any good takes longer, and that part matters more. A PDF can hold clean sentences, pictures of sentences, or six words to a slide, and the questions are only as sound as what was read from it.

EdProcto publishes this guide, and the quiz maker on its [home page](https://edprocto.com/) does the writing step: paste a chapter's text, pick a shape, and Procto AI writes the questions and shows where in your text each answer is.

The short version: one chapter at a time, real text and not an image of it, shapes that fit the content, and five checks on each question before anyone sits it.

## What makes a PDF easy or hard to turn into a quiz?

A PDF is easy to turn into questions when it holds real text in full sentences, and hard when it is a picture of a page, a deck of slides with a few words on each, or mostly tables and formulae.

| What the PDF holds | What goes wrong | What to do |
| --- | --- | --- |
| A chapter with real text | Running heads and page numbers land mid-sentence, and two columns can interleave | Copy one section and delete the debris |
| Lecture slides exported to PDF | Headings and bullets, without the explanation that was spoken | Add your speaker notes, or a paragraph a slide |
| Tables | Cells become a run of words with the rows lost | Rewrite the table as sentences |
| Formulae and symbols | Superscripts, fractions and Greek letters flatten or vanish | Check each against the page |
| A scan or a photograph | No text to read | Have the text recognised first |

### How to tell real text from a scan

Try to select one sentence with the cursor. If the words highlight one at a time, the file has a text layer. If the whole page highlights as a single block, the page is an image.

### Why one chapter at a time

A long document should go in one chapter or section at a time. The reason is the checking: ten questions can be read against twelve pages in a few minutes, and not against three hundred.

## How do you make a quiz from a PDF, step by step?

Get the text out, keep what matters, choose the shapes, have the questions written, and check each one. The first four steps take minutes. The fifth is where the time goes.

1. **Get the text out.** Select the chapter, copy it and paste it into a plain document. Delete running heads, page numbers and the reference list.
2. **Keep what matters.** Cut what you would never test: the anecdote, the acknowledgements, the further reading. Paragraphs you would not test produce questions you would not set.
3. **Choose shapes that fit.** A list of terms and a worked calculation need different questions.
4. **Generate or write.** Paste the text into the quiz maker on EdProcto's [home page](https://edprocto.com/), say whether you are testing yourself or your class, pick the shape and ask for a sample of 3 or 5 questions. Or write them yourself.
5. **Check every question against the PDF.** A practice quiz needs this as much as an exam does: a wrong key teaches the wrong thing.
6. **Try it, then set it.** Sit the quiz yourself, or hand it to a colleague, before the class sees it.

### Which question shape fits which content?

| What the chapter has | A shape that fits |
| --- | --- |
| Terms and definitions | [Fill in the blank](https://edprocto.com/templates/fill_blank), or [matching pairs](https://edprocto.com/templates/matching) for a set |
| A process or procedure | [Sequence the process](https://edprocto.com/templates/process_seq) |
| A classification | [Sort into groups](https://edprocto.com/templates/categorize) |
| An explanation of why | [Multiple choice](https://edprocto.com/templates/mcq), or a [short answer](https://edprocto.com/templates/short_answer) |
| A worked calculation | [Estimate](https://edprocto.com/templates/estimate), a number marked by how close it is |
| A labelled diagram | [Hotspot](https://edprocto.com/templates/hotspot), built by hand from the image |

Each shape has its own page under [templates](https://edprocto.com/templates). The shape menu on the home page leaves out those that need an image to click, a file to upload or a recorded voice.

## How good are automatically generated questions?

Automatically generated questions are a quick first draft and not a finished quiz. In the two studies we read that counted them, 32% and 40% of the questions were fit to use as they stood, and some of the rest were plainly wrong. Where machine questions were compared with people's, the averages were close, though not always equal.

**Fast, and rated close to human.** Cheung et al. (2023) had ChatGPT write 50 multiple-choice questions from medical textbook passages while two professors wrote 50 from the same passages. The machine took just over 20 minutes and the professors about three and a half hours. Five assessors, not told which was which, gave totals that did not differ significantly (37.36 against 38.16 out of 50), though the machine's scores were spread more widely. No students sat the questions.

**A teacher's own text, pasted in.** Ngo et al. (2024) pasted their own immunology study guides into ChatGPT 3.5 and asked for questions with explanations. Of 60 questions, 19 (32%) could be used with minor formatting, 26 (43%) needed substantial changes, most often because explanations they had asked for were missing, and 15 (25%) had answers that were incorrect or seriously misleading. In one, several options were correct and only one was marked as the key.

**Straight from PDFs.** Ayub et al. (2023) loaded eight continuing-education articles into ChatPDF, a tool that passes a PDF to ChatGPT 3.5, and two dermatologists judged the 40 questions against a specialist board examination. Sixteen (40%) were accurate and pitched at the right level. Ten were too simple, nine were vague and five were inaccurate.

**In front of students.** Laupichler et al. (2024) set an experienced educator's questions beside ChatGPT 3.5's in a practice test taken by 161 students. The two sets did not differ significantly in difficulty, but the human questions separated stronger students from weaker ones better (mean discrimination .36 against .24). Bhandari et al. (2024) found no significant difference in difficulty or discrimination between 15 algebra questions ChatGPT wrote from a textbook lesson summary and 15 of the textbook's own, tried on 207 respondents.

A systematic review by Artsi et al. (2024) covered eight studies of machine-written medical questions from 2023. Five judged them valid for examinations, three did not evaluate them, and all eight presented some faulty questions.

### The faults to expect

The same faults recur across these studies:

- a wrong key, or a wrong explanation;
- two or more right options;
- a question so simple that it separates nobody;
- a vague stem;
- wrong options no informed student would pick;
- content that is not in the source, which Cheung et al. (2023) tried to limit by giving the tool a reference text.

### What these studies cannot tell you

These studies tested particular tools on particular material, mostly medicine and mostly a 2023 version of one chatbot. Tools change within months, so a 2023 figure is a reason to check, not a forecast for your chapter. None of them tested EdProcto.

People's questions are not a gold standard either. Tarrant et al. (2009) analysed 514 teacher-written questions from one nursing programme and found only 52.2% of the wrong options doing any work. The checks below apply whoever wrote the question.

## How do you check a generated question?

Check five things against the PDF, and delete any question that fails one of the first three.

| Check | How to do it | If it fails |
| --- | --- | --- |
| Is the answer in the PDF, and where? | Find the sentence and note its page. EdProcto shows it under each question | Delete the question |
| Is the key right? | Answer it yourself with the page open, before you look at the key | Correct the key, or delete |
| Is there exactly one right answer? | Ask whether a well-read student could defend any wrong option | Reword it, or use [all that apply](https://edprocto.com/templates/all_apply) |
| Would the wrong options tempt anyone? | Each should be a mistake students really make | Swap in an error from last year's scripts |
| Does it test the idea, or a sentence? | Ask whether a student who understood, without memorising the wording, would get it | Rewrite it with a new example |

In EdProcto's quiz maker each question carries a line headed "From your material" saying where its answer is, and with the check against the material on, Procto AI is asked to write only questions your text answers. Read each one against the place it names. That removes one fault. It does not prove the key or the wrong options, so read all five. No quiz generator, ours included, gets every generated question right.

The standard list of rules for a multiple-choice question is the taxonomy by Haladyna et al. (2002): 31 guidelines, checked against 27 textbooks and 27 studies. Our guide to [writing multiple-choice questions](https://edprocto.com/guides/how-to-write-multiple-choice-questions-that-measure) goes through them, with most space on wrong options.

## How many questions does a chapter support?

A chapter supports about as many questions as it has ideas worth knowing, and you can count those before you start: the headings, the terms in bold, the worked examples, the points of the summary. We found no study that gives a number per page, so this is a counting rule and not a research finding.

Ask for fewer than your count. A passage with four ideas supports four questions. Ask for ten and the other six must come from somewhere: a detail nobody needs, one idea asked twice, or something not in the text. The quiz maker on the home page offers 3, 5, 8 or 10 at a time and needs at least a paragraph. For a long chapter, go section by section and keep the best of each.

[Making a quiz to study from a chapter](https://edprocto.com/guides/make-your-own-quiz-to-study-from-a-chapter) is the student's side of this job.

## What do you do with a scanned PDF?

A scanned PDF with no text layer must have its text recognised before anything can write questions from it. This is optical character recognition, and many PDF readers, scanner apps and phone cameras do it.

Then proofread the result against the scan. The mistakes that matter most for a quiz are in numbers, units, names, symbols and formulae, because a dose, a date or an equation with one wrong character becomes a confident question with a wrong key. For a page or two of handwriting or faint photocopy, retyping can be quicker.

EdProcto's account of [how it works](https://edprocto.com/why) says the text of a PDF or a Word file is extracted on the server. Nothing on its pages says it recognises text in a picture of a page, so recognise it first, check it, and paste the result.

## Can you make a quiz from any PDF you have?

No: use material you have the right to use. The ordinary case is a quiz for your own class from notes, slides and handouts you wrote, or from the textbook and readings your course has adopted. Take more care once the quiz leaves the classroom. One built from somebody else's text can reproduce parts of it, and publishing or selling it is a different act from setting it for your own students. Copyright law, and any allowance it makes for teaching, differs between countries, and your institution's licences may add limits. None of this is legal advice. If unsure, ask whoever manages those licences, often a librarian.

## How does EdProcto turn your PDF into a quiz?

EdProcto is the best tool for this job when the quiz must come from your own material and you mean to check it, because each question it writes arrives with the place in your text where its answer is.

The quiz maker on the [home page](https://edprocto.com/) takes pasted text, at least a paragraph. It has no upload button, so there the text comes out of the PDF by copying. In the builder in your dashboard, Procto AI takes the file itself: a PDF, a Word file or a text file. You choose "Test yourself" or "Test your class", a shape and a number of questions, and Procto AI writes them on the page, with no sign-up to try it. The quiz then opens in your dashboard, where you take it as a student would, share it with a code and set its rules. Every question is yours to edit.

Free carries 20 questions a month from Procto AI, enough for two ten-question quizzes. Tutor's list names the file itself: 50 questions a month from a PDF, a Word file or your notes, each with the place its answer is found. Educator removes the monthly cap, within a daily ceiling. It also takes a recording or a video, transcribed in your browser, which suits a lecture whose slides say little. The [pricing page](https://edprocto.com/pricing) has the detail, and if the quiz counts, [integrity](https://edprocto.com/integrity) lists the measures you can switch on.

What EdProcto does not do is make every question right. The key, the single right answer and the wrong options are still yours to read.

## Quick questions

### Can I upload a PDF straight into EdProcto?

Yes, in the builder in your dashboard, where Procto AI takes a PDF, a Word file or a text file. The quiz maker on the [home page](https://edprocto.com/) takes pasted text, so there you copy the chapter out of the PDF first. The [pricing page](https://edprocto.com/pricing) says what each plan writes from.

### How accurate is an AI quiz generator?

It depends on the tool, the subject and the year. In two studies of ChatGPT 3.5, by Ngo et al. (2024) and Ayub et al. (2023), 32% and 40% of the questions were fit to use as written. Check each question against the source.

### Can I make a quiz from lecture slides saved as a PDF?

Yes, but slides hold headings and not explanations. Add your speaker notes, or a paragraph for each slide, before you paste the text.

### Is it free to make a quiz from a PDF?

Trying it is. EdProcto's quiz maker writes from pasted text with no sign-up, and the Free plan carries 20 questions a month from Procto AI.

## References

- Artsi, Y., Sorin, V., Konen, E., Glicksberg, B. S., Nadkarni, G., & Klang, E. (2024). Large language models for generating medical examinations: Systematic review. *BMC Medical Education, 24*, Article 354. [https://doi.org/10.1186/s12909-024-05239-y](https://doi.org/10.1186/s12909-024-05239-y)
- Ayub, I., Hamann, D., Hamann, C. R., & Davis, M. J. (2023). Exploring the potential and limitations of Chat Generative Pre-trained Transformer (ChatGPT) in generating board-style dermatology questions: A qualitative analysis. *Cureus, 15*(8), Article e43717. [https://doi.org/10.7759/cureus.43717](https://doi.org/10.7759/cureus.43717)
- Bhandari, S., Liu, Y., Kwak, Y., & Pardos, Z. A. (2024). Evaluating the psychometric properties of ChatGPT-generated questions. *Computers and Education: Artificial Intelligence, 7*, Article 100284. [https://doi.org/10.1016/j.caeai.2024.100284](https://doi.org/10.1016/j.caeai.2024.100284)
- 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 (Hong Kong S.A.R., Singapore, Ireland, and the United Kingdom). *PLOS ONE, 18*(8), Article e0290691. [https://doi.org/10.1371/journal.pone.0290691](https://doi.org/10.1371/journal.pone.0290691)
- Haladyna, T. M., Downing, S. M., & Rodriguez, M. C. (2002). A review of multiple-choice item-writing guidelines for classroom assessment. *Applied Measurement in Education, 15*(3), 309–333. [https://doi.org/10.1207/S15324818AME1503_5](https://doi.org/10.1207/S15324818AME1503_5)
- Laupichler, M. C., Rother, J. F., Grunwald Kadow, I. C., Ahmadi, S., & Raupach, T. (2024). Large language models in medical education: Comparing ChatGPT- to human-generated exam questions. *Academic Medicine, 99*(5), 508–512. [https://doi.org/10.1097/ACM.0000000000005626](https://doi.org/10.1097/ACM.0000000000005626)
- Ngo, A., Gupta, S., Perrine, O., Reddy, R., Ershadi, S., & Remick, D. (2024). ChatGPT 3.5 fails to write appropriate multiple choice practice exam questions. *Academic Pathology, 11*(1), Article 100099. [https://doi.org/10.1016/j.acpath.2023.100099](https://doi.org/10.1016/j.acpath.2023.100099)
- Tarrant, M., Ware, J., & Mohammed, A. M. (2009). An assessment of functioning and non-functioning distractors in multiple-choice questions: A descriptive analysis. *BMC Medical Education, 9*, Article 40. [https://doi.org/10.1186/1472-6920-9-40](https://doi.org/10.1186/1472-6920-9-40)