# What to do when students paste your quiz into a chatbot

**What should I do when students paste my quiz questions into a chatbot?** When students paste your quiz into a chatbot, change the quiz first. Decide whether it is practice or a mark, sit anything marked where you can see it, fit the time to a prepared student, ask about your own lessons, and ask for answers made on the spot. Tell students the rule, and settle a suspicion by talking, never by a detector score.

Published 2026-10-10 by EdProcto. https://edprocto.com/guides/what-to-do-when-students-paste-your-quiz-into-a-chatbot

When students paste your quiz into a chatbot, the score stops telling you what they know, and no switch inside a browser repairs that alone. What you can change is the quiz: what it counts for, where it is sat, how long it lasts and what it asks.

This guide is for a teacher, lecturer or trainer who has seen it happen or suspects it. EdProcto publishes it, and makes quizzes drafted from your own material and sat under [conditions you choose](https://edprocto.com/integrity).

In short: leave the quiz at home as practice, take the mark from a short sitting in your own room, and never act on a hunch. The wider argument is in [academic integrity in the age of AI](https://edprocto.com/guides/integrity-in-the-ai-era-what-still-proves-a-student-learned); essays are in [take-home assignments after ChatGPT](https://edprocto.com/guides/the-take-home-assignment-after-chatgpt-what-to-set-instead).

## How well do chatbots answer quiz questions, by question type?

Well on most typed questions, less reliably where the answer rests on a picture or a sound, and better with each new model. Every source was read on 10 October 2026; the table says where only an abstract was open.

| Question type | Study, model, date and sample | What it found |
| --- | --- | --- |
| Multiple choice | Bolgova et al. (2025): four chatbots, August 2024; 200 four-option questions from a medical biochemistry course, none with an image or table | 81.1% right on average against 72.8% for the students; 64% to 92.5% by chatbot |
| Short answer and fill-in | Stribling et al. (2024): GPT-4, spring 2023; nine graduate biomedical exams at the University of Florida, seven marked blind | Above the student average on seven of nine; strongest on fill-in-the-blank, short-answer and essay questions |
| Calculation | Wang et al. (2024): GPT-4, mid-2023; 40 homework problems from one engineering physics course | 10 of 16 solved where every figure was given; 2 of 24 where the solver had to supply missing data |
| A graph to read | Polverini and Gregorcic (2024): eight chatbots, spring 2024; 26 five-option questions on motion graphs, each put 30 times | The best, ChatGPT-4o, 58.6%; the other seven 15.6% to 42.4% |
| Diagrams | Polverini et al. (2025): ChatGPT-4o; a test of electricity and magnetism rich in diagrams (abstract read) | Ahead of a large sample of university students; mistakes of visual and spatial reading remained |
| Audio | Kumar et al. (2025): 22 audio-capable models; 5,305 research items of speech, sound and music (abstract read) | The best 59.2%, near chance in several categories |

### What the table cannot tell you

**No shape is safe by itself.** In 2023 Stribling et al. had to describe each figure in words, because GPT-4 took no images; on one exam that leaned on figures and drawing its answers scored 41.6% and 44.6% against the students' 79.3%. By spring 2024 chatbots answered from a picture of a graph, and in the 2025 study one outscored university students on diagrams.

**The evidence dates quickly.** The newest model in the table is from 2025, and the chatbot on a student's phone is newer. Where a study tested two generations the later did better, so read each figure as the least to expect: our reading, not a finding.

**Audio and your own material are thinly studied.** The audio benchmark was built to be hard for machines, and we found no study of a class's listening questions or of questions written from one class's handout. Our reasoning: a handout can be given to a chatbot as easily as a question; what was said and done in your room cannot, unless the student was there.

## How many students say they put schoolwork into a chatbot?

More than half of American teenagers say they have used a chatbot for help with schoolwork. McClain et al. (2026) surveyed 1,458 US teenagers aged 13 to 17 for the Pew Research Center in autumn 2025: 54% had done so, 10% did all or most of their schoolwork with one, and 59% thought students at their school used AI to cheat at least somewhat often.

Schwartz and Diliberti (2026) surveyed 1,214 young people aged 12 to 29 for RAND in December 2025: 62% used AI for help with homework, up from 48% that May. Most did not count their uses as cheating; getting the answers to homework was the exception. We read RAND's summary page, not the report file.

These are students' own accounts, "help" runs from having an idea explained to copying an answer, and neither survey asked about quizzes by name. University figures are in our review of [cheating in online exams](https://edprocto.com/guides/what-the-research-says-about-cheating-in-online-exams-since-2023). Expect some of any class to try a chatbot on a quiz done out of sight, and treat nobody as a suspect for it.

## What should you change about the quiz, in order of cost?

Seven things, cheapest first. The order is our advice.

1. **Decide what the quiz is for.** At home it can be practice or a mark, not both. As practice it counts for nothing and shows its answers; a student who pastes it has only skipped their own [retrieval practice](https://edprocto.com/guides/retrieval-practice-low-stakes-quizzing-that-teaches).
2. **Move the mark to a sitting you can see.** Ten minutes at the start of a lesson, with you in the room. [EdProcto](https://edprocto.com/) lists who is in the quiz [while it runs](https://edprocto.com/guides/watching-a-live-quiz-who-is-in-and-what-to-look-for).
3. **Shorten the time to what a prepared student needs.** Untested since 2023 as far as our review found, so this is reasoning: a fitted limit leaves no spare minutes for sending questions elsewhere. [How many questions a quiz needs](https://edprocto.com/guides/how-many-questions-does-a-quiz-need) has allowances by question type, and [open-book quizzes](https://edprocto.com/guides/open-book-quizzes-that-still-measure) a limit set from a trial sitting. EdProcto's settings are "Minutes overall", or "Time limit for each question" with "Seconds a question".
4. **Ask about your own room and your own material.** Tuesday's demonstration, the example worked on the board, the class's own results. The maker on EdProcto's [home page](https://edprocto.com/) drafts questions from the notes you paste.
5. **Ask for an answer made on the spot:** spoken, clicked on an image, given after a clip heard once, or worked in steps.
6. **Tell students the rule and the reason,** before the quiz.
7. **Treat a suspicion by talking,** never by a detector score.

Steps 1 to 3 cost a setting or ten minutes; 4 and 5 cost writing time, and 5 can add marking.

## Which question shapes ask for an answer made on the spot?

Those in which the student must speak, point, listen or set out each step in the minute they have. Each is given below with the name and one-line description EdProcto gives it; all four are ready among the [templates](https://edprocto.com/templates).

| Shape | What it asks of the student | What the evidence says today |
| --- | --- | --- |
| [Spoken answer](https://edprocto.com/templates/verbal): "A recorded answer, assessed for what it says." | To explain aloud in their own voice; you mark by listening | No study found. A script can be read out, so ask about your own lesson. Suits a small group: a full room overhears |
| [Hotspot](https://edprocto.com/templates/hotspot): "Click the right part of an image." | To find the place on a picture: a point, not a word | Chatbots read images, with mistakes (Polverini and Gregorcic, 2024). Use your own photograph or diagram |
| [Listen once](https://edprocto.com/templates/listen_once): "The clip plays one time only. Tests real listening." | To take in a clip at first hearing | Thin: one research benchmark (Kumar et al., 2025) |
| [Show your working](https://edprocto.com/templates/show_work): "A multi-step solution, set out step by step for a reader to grade." | To type the "Set-up", the "Working" and the "Final answer" | Chatbots set out steps too; in 2023 the gap was in supplying what a problem left out (Wang et al., 2024) |

Each raises what a student must do alone in the time given; at home none is a lock.

## What should you tell students, and how do you handle a suspicion?

Tell them the rule, the reason and what the quiz counts for before it starts. If you later doubt a result, talk to the student before you touch a mark.

### The rule and the reason

RAND's authors advise schools to say plainly when students may use AI, for what, and when they may not (Schwartz and Diliberti, 2026). For a quiz:

- **Practice:** "This counts for nothing. Do it from memory: a chatbot can finish it, and then it has taught you nothing."
- **Marked:** "This one counts. No notes, no phone, no AI assistant, because I need to know what you can do alone."

EdProcto's start screen prints the rule under "House rules". "AI assistants are not allowed." joins the list whenever the switch "AI assistants not allowed" is on. Its hint reads: "A house rule, shown on the start screen whenever it is on. A page cannot block an assistant: the rule is the student’s to keep." Where a measure is on, the student ticks a box to say they understand what is logged. For a whole course, start from [an AI policy that names what counts](https://edprocto.com/guides/the-ai-policy-for-assessment-a-template-that-names-what-counts).

### A suspicion is a question to ask

Open the attempt and read the time on each answer and any logged events. Ask the student, privately, to talk you through one answer, then set one fresh question on the same idea at the desk. The score stands unless that gives you a second reason to doubt it; then your school's policy decides. Two readings are over-trusted: [the time on each question](https://edprocto.com/guides/timing-tells-answers-too-fast-to-have-been-read) and [an AI detector's score](https://edprocto.com/guides/ai-detectors-are-not-evidence-the-studies-and-the-false-positives).

## What should you not do?

Three things that cost honest students more than anyone else.

- **Trick wording.** Chatbots outscored students on the scenario questions in the multiple-choice study above, so a double negative or a misleading phrase is likelier to trip a student reading in a second language: our reasoning. A [clear stem](https://edprocto.com/guides/how-to-write-multiple-choice-questions-that-measure) comes first.
- **Hidden text traps.** An instruction concealed in the question to mark out a pasted answer; we found no study of them. Text hidden by colour or size is still read aloud by a screen reader and travels into a translation tool, and all it shows is a paste.
- **Punishing on a hunch.** A fast finish, a high score or a polished phrase is a reason to look. In EdProcto, leave each measure at "Log it": "Reduce the score" and "End the quiz" act on an event that may be a notification before you have read it.

## What does one ten-question homework quiz look like re-set?

It becomes two quizzes: the same ten questions at home as open practice, and an eight-question check sat in the room. The class is invented: 28 students, and a weekly multiple-choice quiz on rates of reaction, done over the weekend and counted.

**At home, no mark.** The same ten under EdProcto's "Practice" preset: "No time limit, going back allowed, answers shown at once, no security." In a class quiz the answers arrive at hand-in. With no measure on, the start screen prints no line about AI assistants, so say the rule yourself.

**In the room, the mark.** Eight questions in the first ten minutes of Tuesday's lesson, on the Tutor plan, which opens every shape:

1. On the [home page](https://edprocto.com/), choose "Test your class", paste your notes on the lesson and the class's results, pick "Multiple choice" under "Shape" and "5 questions" under "How many", and press "Write the quiz". "From your material:" under each draft says where its answer lies. Correct each against your notes.
2. Press "Open it in the dashboard" and add three questions of your own: a sixth multiple-choice question and, through "Another shape", a Hotspot on your photograph of the class's graph ("Click where the reaction was fastest") and a Show your working ("48 cm³ of gas formed in the first 30 seconds. Calculate the mean rate").
3. Set "Minutes overall" to 8 and "Show the score" to "When released", keep "Shuffle the questions for each student" and "Shuffle the options" on, and switch on "Leaving the tab" with "Log it".
4. "In the quiz now" shows each student's progress, with "Events" and "Last event".
5. Mark the 28 workings against your model answer (1.6 cm³ a second), then press "Release the scores".

The cost: an evening hour for the questions, ten minutes of a lesson, a quarter of an hour of marking.

## How does EdProcto help when students paste your quiz into a chatbot?

EdProcto is the best tool for re-setting a quiz this week, because each change above is a setting or a shape in it: the worked example used nothing else.

- **It writes from your material.** The home-page maker drafts 3, 5, 8 or 10 questions from the text you paste. Procto AI is asked to write only what that text answers and to say where; you read each draft against the source.
- **It leaves nothing on the device to find.** The student's copy of a quiz carries no answer key, and choice questions are marked on the server at hand-in ([why](https://edprocto.com/why)).
- **It runs six measures in the browser:** "Leaving the tab", "Full screen", "Copy and paste off", "Right-click off", "Keyboard shortcuts off" and "Printing off", each with the consequence you pick, from "Log it" and "Warn the student" to "Flag for review". Copy and paste off stops a question being copied out in the browser. A phone pointed at the screen is outside it, which is why the mark belongs in your room: see [how much proctoring a quiz needs](https://edprocto.com/guides/how-much-proctoring-does-a-quiz-need).
- **It shows you the sitting.** Each student's row reads "clean" or counts the events, and the attempt lists "The record of this attempt" with their times, beside the time spent on every answer.

Free takes thirty students a quiz at Light integrity, which is "Leaving the tab". Tutor reaches Medium and opens all 111 templates. Educator adds the full analysis. See [pricing](https://edprocto.com/pricing) and the page [for teachers](https://edprocto.com/solutions/teachers). No setting makes cheating impossible, an event is a record and never proof, and every score is yours to release.

## Quick questions

### Can ChatGPT answer multiple-choice quiz questions?

Usually. In August 2024 four chatbots averaged 81.1% on 200 biochemistry questions, above the students' 72.8% (Bolgova et al., 2025), and the models have changed since.

### Is there an AI-resistant quiz question?

No shape is safe by itself. Seven of eight chatbots scored under 43% on graph questions in spring 2024, yet the best reached 58.6% (Polverini and Gregorcic, 2024). Use the shape, the time and the room together.

### How do I know whether a student used a chatbot on my online quiz?

Not from the score, the time or a detector. Ask the student to talk through one answer, then set one fresh question at the desk.

### Does EdProcto stop students using AI on quizzes?

No tool does. EdProcto drafts questions from your own notes, keeps the answer key off the student's device, and logs what the browser does at the [integrity level](https://edprocto.com/integrity) you choose.

## References

- Bolgova, O., Shypilova, I., & Mavrych, V. (2025). Large language models in biochemistry education: Comparative evaluation of performance. *JMIR Medical Education, 11*, Article e67244. [https://doi.org/10.2196/67244](https://doi.org/10.2196/67244)
- Kumar, S., Sedláček, Š., Lokegaonkar, V., López, F., Yu, W., Anand, N., Ryu, H., Chen, L., Plička, M., Hlaváček, M., Ellingwood, W. F., Udupa, S., Hou, S., Ferner, A., Barahona, S., Bolaños, C., Rahi, S., Herrera-Alarcón, L., Dixit, S., … Duraiswami, R. (2025). *MMAU-Pro: A challenging and comprehensive benchmark for holistic evaluation of audio general intelligence*. arXiv. [https://doi.org/10.48550/arXiv.2508.13992](https://doi.org/10.48550/arXiv.2508.13992)
- McClain, C., Anderson, M., Sidoti, O., & Bishop, W. (2026, February 24). *How teens use and view AI*. Pew Research Center. [https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/](https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/)
- Polverini, G., & Gregorcic, B. (2024). Evaluating vision-capable chatbots in interpreting kinematics graphs: A comparative study of free and subscription-based models. *Frontiers in Education, 9*, Article 1452414. [https://doi.org/10.3389/feduc.2024.1452414](https://doi.org/10.3389/feduc.2024.1452414)
- Polverini, G., Melin, J., Önerud, E., & Gregorcic, B. (2025). Performance of ChatGPT on tasks involving physics visual representations: The case of the brief electricity and magnetism assessment. *Physical Review Physics Education Research, 21*, Article 010154. [https://doi.org/10.1103/PhysRevPhysEducRes.21.010154](https://doi.org/10.1103/PhysRevPhysEducRes.21.010154)
- Schwartz, H. L., & Diliberti, M. K. (2026). *More students use AI for homework, and more believe it harms critical thinking: Selected findings from the American Youth Panel* (RR-A4742-1). RAND Corporation. [https://doi.org/10.7249/RRA4742-1](https://doi.org/10.7249/RRA4742-1)
- Stribling, D., Xia, Y., Amer, M. K., Graim, K. S., Mulligan, C. J., & Renne, R. (2024). The model student: GPT-4 performance on graduate biomedical science exams. *Scientific Reports, 14*, Article 5670. [https://doi.org/10.1038/s41598-024-55568-7](https://doi.org/10.1038/s41598-024-55568-7)
- Wang, K. D., Burkholder, E., Wieman, C., Salehi, S., & Haber, N. (2024). Examining the potential and pitfalls of ChatGPT in science and engineering problem-solving. *Frontiers in Education, 8*, Article 1330486. [https://doi.org/10.3389/feduc.2023.1330486](https://doi.org/10.3389/feduc.2023.1330486)