Reducing Cheating in Online Exams Without Full Proctoring
To prevent cheating online exams without a live or recorded proctor, match each control to the cheating method it actually addresses, and scale the effort to the stakes. Proctor-free exam security is the set of design, delivery and review controls that make cheating harder, riskier or less useful when nobody watches the candidate.
Most advice lists a dozen settings and implies they all work equally well. They do not. Some controls have experimental support, some rest on practitioner habit, and one popular control, the honor code, has mixed results. This post maps five cheating methods to controls, grades the evidence as of October 2026, and says when you still need a proctor.
It is written for small training providers. For the buying decision itself, see our guide to proctoring for training providers; this post covers what to do before, or instead of, that purchase.
Key takeaways
- Design controls such as question pools, narrow time windows and one attempt mainly limit answer sharing and item harvesting, not impersonation.
- A clear warning about real detection and consequences has direct experimental support; a signed honor code has weak and contested support.
- Rewording publisher or web-searchable questions cut the score advantage seen on verbatim test bank items in one study, but AI tools mean “Google-proof” is no longer “answer-proof”.
- Full proctoring is still warranted when a result is high-stakes or others rely on it, because only a person or recording can check who sat the exam.
What are the five ways people cheat in online exams?
Online exam cheating falls into five methods: answer sharing (passing answers between candidates over time), collusion (live help during the exam), item harvesting (copying questions to share or sell), lookup (searching notes, websites or AI tools), and impersonation (someone else sits the exam). Each needs a different control, so one setting rarely covers all five.
How common is it? A 2023 systematic review by Newton and Essex, published in the Journal of Academic Ethics, pooled 19 studies with 4,672 participants and found 44.7% self-reported cheating in online exams: 29.9% before COVID-19 and 54.7% during it. The authors judged that most samples probably under-report, and 23 of 25 samples were convenience samples. The review is not specific to training providers, and no comparable figure for adult training providers turned up in the sources fetched for this post, so use these numbers as a signal, not a forecast.
Which design controls prevent cheating online exams, and for which method?
Randomized question pools, narrow time windows and one attempt mainly limit answer sharing and harvesting. Rewritten, application-based items limit lookup. A stated warning reduced cheating in the experiments cited below. None of these confirms who is sitting the exam, so impersonation and live collusion are the gaps that design alone cannot close.
The table below maps each cheating method to the controls that need no proctor and shows what each set leaves uncovered. The pairings are this post’s synthesis of the sources cited in the next section.
| Cheating method | Controls that need no proctor | What stays uncovered |
|---|---|---|
| Answer sharing | Random draws from a larger bank, shuffled options, one start time or a narrow window, one attempt, new items each term | Candidates who sit the exam at the same time |
| Collusion | One item at a time, tight time limit, different forms per candidate, a stated warning, after-exam answer-similarity review | Live help by phone or chat is only caught afterward, if at all |
| Item harvesting | Large bank with rotation, limited answer feedback, delayed results release, retiring exposed items | A harvester who sits the exam legitimately |
| Lookup | Application-based items, reworded publisher items, open-book redesign, lockdown browser for on-device search | A second device or an AI tool on a phone |
| Impersonation | Individual logins, a scheduled session with a staff identity check, comparing exam results with in-course performance | Cannot be closed without a person or recording verifying identity |
Pool size is easy to get wrong, so here is the arithmetic (this post’s calculation, assuming each candidate gets an independent random draw and no exposure control). Two candidates who each get 30 items from a bank of 60 share about 15 items on average, half the exam. From 90 items they share about 10, and from 150 about 6. Because a learner may sit the exam more than once, attempt limits matter too: after three attempts a candidate has seen about 88% of a 60-item bank, 70% of a 90-item bank and 49% of a 150-item bank. Our guide to building an item bank for a training company covers usage counts and retirement rules.
How strong is the evidence for each control?
Evidence is strongest for detection warnings and for rewording lookup-prone items, both tested directly. It is weak or contested for honor codes and open-book redesign, and mostly practitioner opinion for randomization, time limits and attempt limits. Lockdown browsers have only a partial effect. Read the table as a ranking of confidence, not of importance.
| Control | What the sources found | Confidence |
|---|---|---|
| Warning with stated consequences | Corrigan-Gibbs and colleagues (ACM TOCHI, 2015) measured cheating directly in two experiments (632 exam takers, 2,378 task participants); a stern warning roughly halved it. Pleasants and colleagues (Online Learning, 2022) found a warning about detection technology and penalties cut page-leaving to 15% of students. | Moderate: direct measurement, limited settings |
| Honor code or pledge | The 2015 experiments found a small, non-significant effect, and Pleasants found appeals and pledges did not reduce cheating. A 2018 registered replication (25 labs, 4,647 participants) of the Mazar, Amir and Ariely moral-reminder effect found none. McCabe and colleagues’ 2001 review says honor codes can influence behavior, with context such as peer behavior the strongest factor. | Contested; cheap, so use it, but do not rely on it |
| Reworded or application items | Golden and Kohlbeck (Journal of Accounting Education, 2020) used verbatim and paraphrased test bank questions in three online courses: 80.4% correct on verbatim, 69.1% on paraphrased, a gap that held after controlling for an honor code statement and proctoring technology. | Moderate: one subject, one institution |
| Pools, shuffling, one item at a time, one attempt, narrow window | Cluskey, Ehlen and Raiborn proposed eight such procedures in a position paper with no sample or comparison group. The 2022 ITC and ATP guidelines recommend multiple fixed forms to manage exposure. No head-to-head test of these settings turned up in the sources fetched. | Reasoned practice, not measured |
| Strict time limit | Richter and colleagues (Frontiers in Veterinary Science, 2024) suggest a strict limit may help in unsupervised exams, based on in-person response-time logs, not a cheating test. They warn it can disadvantage some learners. | Weak, with an accessibility cost |
| Lockdown browser | The ITC and ATP guidelines state: “Locked-down browsers create barriers but do not prevent all cheating and content theft methods.” | Partial barrier |
| Open-book redesign | A review of 37 studies (Durning and colleagues, Academic Medicine, 2016) found no sufficient evidence to use either format exclusively; studies of exam performance favored closed-book scores. None of the sources fetched for this post measured cheating. | No cheating evidence either way |
The ITC and ATP guidelines in that table are the Guidelines for Technology-Based Assessment, released in November 2022.
Does an unproctored online exam still measure learning?
Often yes, but the evidence is contested. Chan and Ahn (PNAS, 2023) studied nearly 2,000 students at one Midwestern university and found unsupervised online exam scores correlated strongly with in-person scores, with no sign that weaker students gained more. Newton critiqued the work; Chan and Ahn’s reply reports average online score inflation of about 5% (g = 0.40) alongside r = 0.59. The study predates wide generative AI use.
That matters: a 2024 PLOS ONE study (Scarfe and colleagues) entered fully GPT-4-written answers into five undergraduate psychology modules at one UK university, and about 94% went undetected. It does not test your item types, but a scenario-based question that needs judgment about a specific situation is still harder to paste into a search box than a recall question, though not impossible for an AI tool.
Pro tip: Only Warn About Controls You Run. The warnings that worked in the studies spelled out consequences, and in one case detection technology. Make that statement true: if you tell learners that answer patterns are reviewed, schedule the review. An empty threat costs you credibility the first time someone tests it.
What can you do to prevent cheating online exams at low, medium and high stakes?
Low-stakes quizzes need design controls and a warning, nothing more. Medium-stakes course exams add pools, a narrow window, one attempt, application items and after-exam review. High-stakes exams that others rely on, such as a credential, need supervised delivery as well, because design controls cannot verify identity.
The table shows one way to tier controls. These tiers are this post’s own framing, not a standard; the ITC and ATP guidelines say only that the level of supervision depends on the testing stakes and context. To score a specific exam, use the four-question method in the proctoring guide linked above.
| Stakes tier | Typical use | Controls | Is full proctoring needed? |
|---|---|---|---|
| Low | Knowledge checks, practice tests, pre-course diagnostics | Low point value, random draw, immediate feedback, a short integrity notice | No |
| Medium | End-of-course exam for a completion certificate | Pool sized to the overlap you accept, shuffled options, narrow window, one attempt, application items, stated review, after-exam flags | Usually no; consider supervising the final session |
| High | Credential, licensing prerequisite or safety-critical sign-off that third parties rely on | All medium controls plus form rotation and item retirement | Yes: in-person invigilation or remote proctoring |
How can you catch cheating after the exam without a proctor?
Look for patterns the exam data can show: unusually similar answers between candidates, implausibly fast or slow responses, items that get easier over time, and unusual cohort pass rates. Treat each flag as a clue for human review, never as proof. A flag should start a conversation or a supervised retest, not an automatic penalty.
The Handbook of Quantitative Methods for Detecting Cheating on Tests (Cizek and Wollack, 2017) groups methods into three threats: similarity and answer copying, preknowledge and item compromise (including response-time analysis), and test tampering or unusual score gains. The Federation of State Boards of Physical Therapy describes similar analyses (answer similarity, response patterns, response time and item drift) in its licensing exam program and states that an aberrant pattern “is not necessarily an indication of cheating.”
Scale is the practical catch. Published validation often uses large data: Vista (2019) tested the omega index, an item response theory answer-copying index, on simulated cheating in PIRLS 2011 data from 64,232 test takers, and it caught 16 of 18 simulated groups. Do not assume the same power for a cohort of 15. For small cohorts, simpler checks work better: identical wrong answers on the same items, completion times far outside the group, and a learner who aces the exam but struggled in class. Our overview of AI-driven cheat detection in online assessments describes the general idea of automated flagging; the point here is that flags need a human decision behind them.
Pro tip: Write the Flag Rule Before the Exam. Decide in advance which pattern triggers a review, who reviews it, and what the outcome is, such as a short supervised retest. Deciding after you see a name invites bias and weakens any later dispute.
When is full proctoring still warranted?
Full proctoring is warranted when a result is high-stakes, when impersonation or live outside help is a credible threat, or when third parties rely on the result as proof of competence. Design controls reduce shared answers and lookup, but they cannot confirm who sat the exam or who else was in the room.
The comparison studies point the same direction, with caveats. In the Newton and Essex review only two studies compared proctored and unproctored exams; both found less self-reported cheating with proctoring, including web searching reported by 39.8% of unproctored versus 8.5% of proctored students. Dendir and Maxwell (Computers in Human Behavior Reports, 2020) saw average scores fall in two online courses once webcam proctoring began, and Alessio and colleagues (Online Learning, 2017) found proctored students scored about 17 points lower. Lower scores are consistent with less cheating but do not prove it. Our proctoring guide covers a 2025 experiment with a different result.
A middle path suits many providers: apply design controls to every exam, and supervise only the one that carries the pass or fail decision, in person or at a partner site. Proctoring also does not fix leaked items or weak pass marks, so the design controls stay in place either way.
What should you do next?
List your exams and put each in a tier from the table. Start with the medium-stakes exams, because they gain the most from design controls: build or expand the pool, reword any publisher or web-searchable items into application items, and add a warning that describes only what you will actually do.
Then check what your platform supports before you design around it. As of October 2026, the SimpliAssess features page lists item banking, schedule management and analytics and reporting; it does not describe randomization, time limits, attempt limits or proctoring, so confirm each control with any vendor in writing.
Frequently asked questions
Can you prevent cheating in online exams without proctoring?
You can reduce it, not eliminate it. Question pools, narrow windows, one attempt, application-based items and a warning about real consequences lower answer sharing and lookup. They cannot verify identity or stop live help on a second device. For credentials others rely on, add in-person invigilation or proctoring.
Does randomizing questions stop cheating?
Randomization reduces answer sharing but does not stop it. Two candidates who each get 30 items from a 60-item bank still share about 15 on average. A larger bank, shuffled options and one attempt lower the overlap. The sources fetched include no experiment on randomization alone, so treat it as reasoned practice.
Do honor codes reduce cheating on online exams?
The evidence is mixed. Two direct online experiments found small, non-significant effects from honor codes, while warnings about consequences roughly halved cheating in one. A 2018 registered replication found no moral-reminder effect. Institution-wide integrity cultures may matter more than a pledge. Use a code because it is cheap, not because it is proven.
Are open-book exams less prone to cheating?
Open-book exams change what counts as cheating, because looking up information is allowed. They do not stop collusion, answer sharing or impersonation. A review of 37 studies found no sufficient evidence to use open-book or closed-book exclusively, and none of the sources fetched for this post measured cheating, so redesign items for application, not recall.
Does a lockdown browser stop cheating?
Only partly. The ITC and ATP technology-based assessment guidelines say locked-down browsers create barriers but do not prevent all cheating and content theft methods. A lockdown browser limits on-device searching, not a second phone, a helper in the room or someone else taking the exam.
How do you detect cheating after an online exam?
Review answer similarity between candidates, unusual completion times, items that become easier over time, and results that do not match in-course performance. Treat each flag as a clue, not proof. Statistical methods were mostly validated on large datasets, so with small cohorts rely on simple checks and a human decision.
How many questions should an exam pool have?
No standard number exists, so size it by the overlap you accept. With 30 items drawn per exam, a 60-item pool gives about 15 shared items between two candidates, 90 gives about 10 and 150 gives about 6. Our proctoring guide suggests two to three times as many items as one exam uses; that is its own suggestion, not a standard.
