Cheating in online exams has been studied closely since 2023, and the studies disagree about how much of it there is. One review found nearly half of students admitting it; one large survey found under 3%. Both figures are honest; they measure different things.
This review, for a teacher or an exams officer, sets the studies side by side: what each measured, on whom, and what it found. EdProcto publishes it and makes quizzes sat under conditions you set. What still counts as proof is argued in academic integrity in the age of AI; this is the research around it.
Every source was read on 10 October 2026, and the table says where only an abstract was open.
What did the main studies since 2023 find?
Admitted cheating anywhere from 3% to a half, a chatbot that answered nine in ten medical exam questions, small score gaps between watched and unwatched sittings, and real costs to students who are monitored or accused.
| Study | Design and sample | Main finding |
|---|---|---|
| Newton and Essex (2024) | Systematic review of 19 survey studies: 4,672 university students, searched in mid-2022 | 44.7% admitted cheating in an online exam: 29.9% before the pandemic, 54.7% during |
| Henderson et al. (2023) | Anonymous survey straight after online exams, one Australian university, 2021: 7,511 complete replies from 39,308 students | 2.9% said they had cheated; of the rest, 18.9% had been tempted |
| Ibrahim et al. (2023) | ChatGPT's answers in 32 courses at one university, graded blind beside students', January 2023 | Level with or ahead of students in nine courses; behind in mathematics and economics |
| Newton et al. (2025) | ChatGPT-4o on medical multiple-choice questions, mid-2024; 62 newly written | 60 of the 62 new questions right; 2 of 14 where the options were labels on a picture |
| von Garrel and Mayer (2023) | Online survey, 6,311 students in Germany, mid-2023; not a random sample | 63.4% had used AI tools for their studies |
| Gulati et al. (2024) | Two randomised crossover experiments, six exams, one computer science class (abstract read, no headcount) | Unwatched scores 2.3% higher than on Zoom for non-coding questions; 5.4% where all met the same version |
| Oeding et al. (2024) | Course grades before and after Proctorio: 252 students, six online courses, one US university | No significant difference overall; lower grades in one course |
| Zhao et al. (2024) | Four randomised experiments in classroom exams: 692 students, one Chinese university | Less cheating with no invigilator present: 7.9% against 22.7% in the first |
| Mizori et al. (2024) | Meta-analysis of eight open-book against closed-book comparisons, all before chatbots | Open-book scores 5.91 points higher, with results far apart |
| Pokorny et al. (2023) | Survey: 375 students, 11 science courses, three US universities, spring 2021 | 68.3% of those who ranked their concerns were very concerned about a wrongful flag |
| Gorichanaz (2023) | 49 Reddit threads by students accused of using ChatGPT (abstract read) | Most said the accusation was false; no rate follows |
How common is cheating in online exams?
By students' own account, anywhere from 3% to a half, depending on who was asked and about what. In the review by Newton and Essex (2024), cheating alone was more common than cheating with others, and the reason given most often was that the opportunity was there. Its searches were run in 2022, so every survey in it predates public chatbots.
Henderson et al. (2023) asked a narrower question, about the exams a student had just sat, and the answer was far lower.
Why do self-reported figures and other measures disagree?
Because an admission depends on who volunteers and what they fear. Of the review's 25 samples, 23 were convenience samples, and none used a randomised-response method or an incentive to answer truthfully; the authors expect under-reporting. In the Australian survey about four students in five did not complete it.
A direct count lands in between, though not online. Zhao et al. (2024) planted questions on invented concepts in real classroom exams and posted a made-up key online as each exam began, so a right answer could only have been looked up. Between 3.5% and 26.3% of students gave one, depending on the experiment and the condition. We found no study since 2023 that measured the rate directly in an online exam.
What changed with chatbots?
A recent chatbot answers most multiple-choice exam questions put to it, and most students surveyed say they use one to study. Whether cheating in online exams rose as a result has not been measured.
How well do chatbots answer exam questions in different subjects?
Well in most subjects tested, and better with each model. In January 2023 ChatGPT matched or beat students in nine of 32 courses, with at least one in every discipline except mathematics and economics (Ibrahim et al., 2023). In mid-2024 ChatGPT-4o scored 94% on UK medical licensing questions and answered 60 of 62 that the authors had newly written, so it was not repeating answers it had seen (Newton et al., 2025).
Where the options were labels on a picture, the model got 2 of 14 right, against 13 of 14 for the same questions in words. That is one model in one year, and the authors conclude that secure testing environments are required.
How often do students report using chatbots?
Most report using them to study; about a fifth, in one survey, to cheat. In von Garrel and Mayer (2023), 63.4% had used AI tools for their studies, most often to have a concept explained. In the five-country survey by Ibrahim et al. (2023), 74% of students said they would use ChatGPT.
Times Higher Education reported that 22% of the students surveyed for a 2025 paper by Newton admitted using AI to cheat in an assessment in the previous twelve months, and that 55% had been assessed by unsupervised online exams (Rowsell, 2025). We could not open the paper itself.
What did proctoring change when proctored and unproctored sittings were compared?
It moved scores a little; what it does to cheating itself rests on a few surveys.
Gulati et al. (2024) ran the cleanest test: the same students sat some exams unwatched and some on Zoom, in an order set by chance. Unwatched scores were higher on non-coding questions only, and a second experiment found no significant difference between Zoom and an invigilated room. Oeding et al. (2024) compared the same courses under the same professors before and after Proctorio recorded webcam, screen and sound; that comparison was not randomised and ran into spring 2020.
Two surveys in Newton and Essex (2024) suggest less cheating under remote proctoring, on data the authors call limited. In Henderson et al. (2023) the least tempted students were those checked in by a human proctor, which the authors suggest raised the sense of being watched.
Zhao et al. (2024) cut the other way. In each of four experiments, though not significantly in one, fewer students looked up the planted answers when no invigilator was in the room, and a reminder about integrity lowered cheating further. One university, in person; still, a warning against assuming that watching always deters.
Do question pools, shuffling, time limits and open-book design reduce cheating?
Different versions of a question have evidence behind them since 2023; the other three have little.
- Question pools. In Gulati et al. (2024) most non-coding questions came in randomised versions. On the few every student met in the same version, the unwatched advantage more than doubled.
- Shuffling the order. Not tested alone since 2023; the lecturer's midterm has the earlier experiments, where order alone still leaked.
- Time limits. We found no study since 2023 that tested one as a defence.
- Open-book design. The comparisons Mizori et al. (2024) pooled, from university science and medicine, differed so much (I² = 97%) that the average is a weak guide. See open-book quizzes that still measure.
None of the four stops a chatbot answering whichever version a student holds: our reasoning, not a finding.
Who is harmed by heavy monitoring and by false accusations?
Honest students: those in poor rooms, those who fear a wrongful flag, and those accused on thin evidence.
In the courses Pokorny et al. (2023) surveyed, which used Proctorio or Respondus LockDown Browser, over half the students (53.4%) sat their exams in poor conditions, such as noise or a weak connection, and 21.4% described emotional distress when using the software.
The accused students Gorichanaz (2023) read wrote of building a case with evidence, and of trust in technology set against trust in students. Forum posts choose themselves, and we found no study of how often such an accusation is wrong. For flags that fall unevenly by skin tone, and anxiety under a webcam, see how much proctoring a quiz needs.
How should you read a study of cheating in online exams?
Ask three things before you repeat its number.
Is it self-reported?
Then it is probably an undercount: a student who cheated is less likely to reply, or to say so. Check whether it asked about one exam just sat or every exam so far.
Who answered?
A convenience sample is whoever was willing. von Garrel and Mayer (2023) say that theirs, reached through course coordinators, is not representative.
When was it gathered?
Most figures come from the emergency remote teaching of 2020 and 2021, when exams moved online within weeks. Henderson et al. (2023) and Pokorny et al. (2023) both list that as a limit. Chatbot results age faster: the 2023 and 2024 scores above came from different models.
Where is the evidence thin or conflicting?
In five places.
- The rate since chatbots. No direct measure in online exams.
- Cameras against chatbots. No comparison we found measured chatbot use.
- Supervision. A small effect in one class, the reverse in four classroom experiments.
- Time limits and shuffled order. Untested since 2023.
- Schools. Every study in the table is of universities.
What can a teacher reasonably conclude?
Five things, none needing an exact rate.
- Some students cheat when the chance is there. Plan for it without treating a class as suspects.
- A question answered out of sight can be put to a chatbot. Put results that decide something under conditions you know: a supervised room, or a live sitting that leaves a record, such as a quiz in EdProcto.
- Give students different versions, or at least a different order. Versions are the one design measure tested since 2023.
- Monitor lightly and say what is on. The score gaps were small and the worries widespread.
- Treat a flag, a fast finish or a high score as a question to ask. No study here supports treating one as proof.
Where does a quiz under known conditions fit?
Where the evidence is firmest: a sitting whose conditions you chose and can state afterwards. EdProcto is the best tool for that sitting, because its settings follow those findings.
- An order of their own. The unwatched advantage shrank where students met different versions. EdProcto varies the order: "Shuffle the questions for each student" and "Shuffle the options" give every student their own.
- Nothing on the device to find. Opportunity was the reason students most often gave. The copy a student opens carries no answer key, and the marks are made on the server (why it is built this way).
- Measures in the browser alone. Most proctored students in one survey feared a wrongful flag. The Exam preset switches on six measures that act in the browser: "Full screen", "Copy and paste off", "Right-click off", "Keyboard shortcuts off", "Printing off" and "Leaving the tab". That is the Medium level. Each carries the consequence you pick ("Log it", "Warn the student", "Reduce the score", "End the quiz" or "Flag for review"); kept at "Log it", an event is a line for you to read.
- A person watching. Students checked in by a person reported less temptation. "In the quiz now" lists each student with "Events" and "Last event", and where the quiz asks for an email address and limits retakes, a second browser is refused until you press "Let back in".
With the Exam preset, scores stay held back until you press "Release the scores", so there is time to ask first. Free saves a quiz at Light integrity and Tutor at Medium, which the Exam preset needs: see pricing. Where a result must be certain, use EdProcto as the paper in a supervised room. No setting makes cheating impossible, an event is never proof, and a score is never a grade by itself.
Quick questions
How common is cheating in online exams?
Nobody knows exactly. Pooled surveys put admitted cheating at 44.7% (Newton and Essex, 2024) and one large survey at 2.9% (Henderson et al., 2023).
Can ChatGPT pass an online exam?
Often. ChatGPT-4o answered 60 of 62 newly written medical questions in 2024 (Newton et al., 2025), and early ChatGPT matched students in nine of 32 courses (Ibrahim et al., 2023).
Does proctoring stop cheating in online exams?
It has not been shown to. Unwatched scores were 2.3% higher than Zoom-proctored ones in one class (Gulati et al., 2024); on cheating itself, online, there are only surveys.
Does EdProcto stop cheating in online exams?
No tool does. EdProcto gives each student an order of their own, keeps the key on the server and logs what the browser does at the integrity level you choose.
References
- Gorichanaz, T. (2023). Accused: How students respond to allegations of using ChatGPT on assessments. arXiv. https://doi.org/10.48550/arXiv.2308.16374
- Gulati, R., West, M., Zilles, C., & Silva, M. (2024). Comparing the security of three proctoring regimens for bring-your-own-device exams. In Proceedings of the 55th ACM Technical Symposium on Computer Science Education V. 1. Association for Computing Machinery. https://doi.org/10.1145/3626252.3630809
- Henderson, M., Chung, J., Awdry, R., Ashford, C., Bryant, M., Mundy, M., & Ryan, K. (2023). The temptation to cheat in online exams: Moving beyond the binary discourse of cheating and not cheating. International Journal for Educational Integrity, 19, Article 21. https://doi.org/10.1007/s40979-023-00143-2
- Ibrahim, H., Liu, F., Asim, R., Battu, B., Benabderrahmane, S., Alhafni, B., Adnan, W., Alhanai, T., AlShebli, B., Baghdadi, R., Bélanger, J. J., Beretta, E., Celik, K., Chaqfeh, M., Daqaq, M. F., El Bernoussi, Z., Fougnie, D., Garcia de Soto, B., Gandolfi, A., … Zaki, Y. (2023). Perception, performance, and detectability of conversational artificial intelligence across 32 university courses. Scientific Reports, 13, Article 12187. https://doi.org/10.1038/s41598-023-38964-3
- Mizori, R., Sadiq, M., Ahmad, M. T., Siu, A., Ahmad, R. R., Yang, Z., Oram, H., & Galloway, J. (2024). STEM exam performance: Open- versus closed-book methods in the large language model era. The Clinical Teacher, 22(1), Article e13839. https://doi.org/10.1111/tct.13839
- Newton, P. M., & Essex, K. (2024). How common is cheating in online exams and did it increase during the COVID-19 pandemic? A systematic review. Journal of Academic Ethics, 22(2), 323–343. https://doi.org/10.1007/s10805-023-09485-5
- Newton, P. M., Summers, C. J., Zaheer, U., Xiromeriti, M., Stokes, J. R., Bhangu, J. S., Roome, E. G., Roberts-Phillips, A., Mazaheri-Asadi, D., Jones, C. D., Hughes, S., Gilbert, D., Jones, E., Essex, K., Ellis, E. C., Davey, R., Cox, A. A., & Bassett, J. A. (2025). Can ChatGPT-4o really pass medical science exams? A pragmatic analysis using novel questions. Medical Science Educator, 35(2), 721–729. https://doi.org/10.1007/s40670-025-02293-z
- Oeding, J., Gunn, T., & Seitz, J. (2024). The mixed-bag impact of online proctoring software in undergraduate courses. Open Praxis, 16(1), 82–93. https://doi.org/10.55982/openpraxis.16.1.585
- Pokorny, A., Ballen, C. J., Drake, A. G., Driessen, E. P., Fagbodun, S., Gibbens, B., Henning, J. A., McCoy, S. J., Thompson, S. K., Willis, C. G., & Lane, A. K. (2023). "Out of my control": Science undergraduates report mental health concerns and inconsistent conditions when using remote proctoring software. International Journal for Educational Integrity, 19, Article 22. https://doi.org/10.1007/s40979-023-00141-4
- Rowsell, J. (2025, June 27). Universities need to "redefine cheating" in age of AI. Times Higher Education. https://www.timeshighereducation.com/news/universities-need-redefine-cheating-age-ai
- von Garrel, J., & Mayer, J. (2023). Artificial intelligence in studies: Use of ChatGPT and AI-based tools among students in Germany. Humanities and Social Sciences Communications, 10, Article 799. https://doi.org/10.1057/s41599-023-02304-7
- Zhao, L., Peng, J., Ke, S., & Lee, K. (2024). Effectiveness of unproctored vs. teacher-proctored exams in reducing students' cheating: A double-blind randomized controlled field experimental study. Educational Psychology Review, 36, Article 126. https://doi.org/10.1007/s10648-024-09965-z