# How to prevent cheating in online exams: what actually works > Twelve measures against cheating in online exams, ranked from the least to the most costly for the honest candidate: what each one prevents, what it does not, and where the GDPR draws the line. No measure makes cheating impossible; a few, combined and chosen by the stakes of the exam, make it costly, uncertain and traceable. Source: https://evalmee.com/en/blog/how-to-prevent-cheating-in-online-exams/ Published: 2026-09-04 Updated: 2026-09-07 ES: https://evalmee.com/es/blog/como-evitar-trampas-en-los-examenes-en-linea/ ES-ES: https://evalmee.com/es-es/blog/como-evitar-que-copien-en-los-examenes-online/ FR: https://evalmee.com/fr/blog/empecher-la-triche-aux-examens-en-ligne/ IT: https://evalmee.com/it/blog/come-impedire-di-copiare-agli-esami-online/ --- No measure makes cheating in an online exam impossible. That is not a reason to drop the question; it is a reason to ask it properly: which measures prevent the most, at the least cost to the candidate who was never going to cheat? This article ranks twelve measures on that scale, from the cheapest to the most intrusive. For each one: what it prevents, what it does not prevent, and, where personal data is involved, where the GDPR draws the line for [remote proctoring](/en/assessment-glossary/#remote-proctoring). Four principles carry most of that line, and they hold anywhere the GDPR applies: proportionality to the stakes of the exam, with no data collected beyond what those stakes justify ([article 5](https://eur-lex.europa.eu/eli/reg/2016/679/oj)); prior information of candidates (article 13); no decision with a legal or similarly significant effect taken by a machine alone (article 22); and, for a system of this kind, a data protection impact assessment run before it goes live (article 35). “Cost to the honest candidate” means everything a measure takes from someone who plays fair: stress, personal data, a quiet room, a computer that meets the requirements. ## 1. Rules announced clearly, in advance Cost: none. A rule known before the exam, on allowed materials, the time limit, what is recorded and what happens when a candidate leaves the exam window, removes the honest candidate’s uncertainty and the cheat’s excuse. It also lets a candidate choose the in-person alternative when one is offered. What it prevents: the grey zone, the “I did not know the notes were forbidden”, and most disputes after the exam. What it does not prevent: anyone who has decided to cheat. It changes what counts as cheating, not whether it happens. The data protection line: this is the duty to inform, in exam form. Article 13 of the GDPR requires that a candidate be told, before the processing begins, what is collected, why, on what legal basis, for how long it is kept and who has access to it. Published early enough, that information also lets a candidate choose the in-person alternative with full knowledge; restated just before they connect, it removes the surprise. Everything below assumes this measure is in place. ## 2. Questions that cannot be looked up Cost: none to the candidate; the cost falls on the author of the paper. An application question, a short answer that has to be justified, a spreadsheet or a diagram to produce, a file to upload: the answer is not in the course notes, not in a search engine, and not in a chatbot without the candidate understanding it first. What it prevents: the most common form of cheating in a home exam, looking the answer up, and most of the value of a leaked paper. What it does not prevent: help from a person who knows the subject, or a generative AI used with skill. The design of the paper, essay or short answer, identical questions or not, time per question, is a lever against cheating as much as a pedagogical choice, and a paper that cannot be looked up may need no proctoring at all. That is data minimisation in its purest form: the measure that collects nothing (article 5). Nothing to declare under the GDPR: no data is collected. ## 3. Anonymous grading and question-by-question grading Cost: none. These two measures do not stop cheating during the exam; they stop the result from being bent afterwards. [Anonymous marking](/en/assessment-glossary/#anonymous-marking) removes the name from the copy. Question-by-question grading, where the grader reads every answer to question 3 before opening question 4, reduces the drift in severity between the first copy and the last. What it prevents: favouritism, unconscious bias, and any claim that the grade was personal. What it does not prevent: anything that happens during the exam. The data protection line: a [data minimisation](/en/assessment-glossary/#data-minimisation) choice. The grader does not need the name to grade. Both modes exist in Evalmee; the [grading page](/en/product/grade/) describes them. ## 4. A paper that cannot be downloaded before the start Cost: none. If the questions are loaded at the moment each candidate opens the exam, and not before, the paper cannot circulate in the hours before the session. What it prevents: a leak from an early download, and the paper being shared with the group that sits later. What it does not prevent: a candidate photographing their screen once the exam has started and sending it to a later session. Several sessions on the same paper call for measure 5. ## 5. Random draw from a question bank and shuffled order Cost: low, provided the drawn papers are of equal difficulty. That is the author’s job, and a fairness problem when it is skipped. What it prevents: two neighbours, or two friends on a call, facing the same question at the same moment; the value of a screenshot sent to a later session; the answer key in the group chat, because “A, C, B, D” means nothing once the choices are shuffled. What it does not prevent: a candidate who has help at hand, or a full leak when the bank is small. Shuffling the choices only helps for closed questions. Nothing to declare under the GDPR. In Evalmee, a random draw pulls the number of questions you set from a [question bank](/en/assessment-glossary/#question-bank), and the order of sections, pages, questions and choices can be randomised; see the [exam creation page](/en/product/create/). ## 6. Time-limited sessions Cost: low to moderate. Time pressure is felt by everyone. A candidate with an access arrangement needs their extra time applied without asking, and a connection incident needs a way to add time. What it prevents: leisurely lookup. A candidate with forty questions in forty minutes has no time to search each one, and the window in which outside help is useful shrinks. What it does not prevent: help already sitting next to the candidate. And a limit set too tight tests speed rather than knowledge. In Evalmee, extra time is attached to the participant’s profile and time can be added during the exam; see the [exam delivery page](/en/product/share/). ## 7. Full-screen mode and screen-exit tracking Cost: moderate. The candidate composes in full screen, and leaving it triggers a warning. Honest candidates do leave the window: a notification, a wrong click, a tab that crashes. What matters is that the exit is recorded, with its count and duration, and never sanctioned on its own. What it prevents: the casual switch to a tab with the notes or a chatbot, or at least makes it visible. What it does not prevent: a phone on the desk, a second screen (measure 9), a person in the room. The data protection line: detecting, or blocking, access to other tabs stays proportionate for an exam whose stakes justify closer monitoring, and it says nothing about the candidate’s body. The limit is that a detected exit is information for the examiner, not a sanction: drawing a consequence from it automatically would be a decision taken by a machine alone, which article 22 of the GDPR restricts. ## 8. Copy-and-paste detection Cost: low. It takes almost nothing from a candidate who types their answers. What it prevents: pasting a block of text from a document or a chatbot into an open answer, or rather makes it visible with a timestamp. Combined with a time limit, it makes the trick expensive. What it does not prevent: retyping, or a short answer read off a second device. It says nothing about where the text came from. The data protection line: this is the candidate’s activity within the exam, not their behaviour. Same rule: an event to review, never a verdict. ## 9. Second-screen detection Cost: low for most candidates. It is a real cost for the one who legitimately works with two screens, for accessibility or because their laptop sits on a dock: announce the rule in advance and decide beforehand how you handle exceptions. What it prevents: the notes or the chatbot on the second monitor, outside the exam window, which is the classic set-up. What it does not prevent: a phone, a tablet or a second laptop, which the exam browser cannot see. The data protection line: the detection is technical, the presence of an additional display, and has nothing biometric about it. It still needs to be announced (measure 1). ## 10. Screenshots on events Cost: moderate. A capture of the candidate’s screen is a capture of a private device. Taking it only when something happens, leaving the window, switching to another application, leaving full screen, rather than continuously, is what keeps it proportionate. It also restricts the exam to computers. What it prevents: the dispute over “I left the window by accident”. The capture shows what was on the screen at that moment, and gives the examiner a fact to read instead of a count. What it does not prevent: anything off screen. The data protection line: data minimisation applied to screen monitoring (article 5), with a [retention period](/en/assessment-glossary/#data-retention-period) set before the system goes live and access restricted to the people who need it. ## 11. Identity check by a human before the exam Cost: moderate, and one-off. The candidate shows an identity document and their face to the webcam before the exam starts, and a person compares the two. What it prevents: impersonation, a friend sitting the exam in the candidate’s place, which no in-exam detection catches. What it does not prevent: anything after the check. An identity verified at 9:00 says nothing about who is at the keyboard at 9:40. That is what measure 12 is for, and even it is imperfect. The data protection line: a document check carried out by a person is the simplest route. Automated comparison of a face with the document processes biometric data under article 9 of the GDPR: it needs a specific legal ground, an alternative always available to the candidate, and it must never build a database of biometric templates. A human comparison sidesteps all of that. ## 12. Webcam and microphone recording Cost: high. The most intrusive measure on the list: a camera and a microphone on, in a private room, for the whole exam. It captures the household as much as the candidate, and it requires a quiet room and a working webcam, which not every candidate has. What it prevents: a person in the room, answers read aloud, a candidate leaving their seat, or rather makes them reviewable afterwards. Its deterrent effect is probably its main effect. What it does not prevent: what happens out of frame, an earpiece, a screen placed behind the camera. Monitoring a private device in a private room stays necessarily imperfect, if only because the camera cannot cover the whole room. The data protection line: proportionate to an exam with high stakes, an entrance exam for instance, and not to a mock exam, which should run without remote proctoring. Candidates are told that the system may capture their surroundings and advised to sit in a quiet, neutral room (article 13). The retention period is set before go-live, access is restricted and logged (article 5), and the recording never feeds an automated analysis of the candidate’s behaviour. A system of this reach is the textbook case for the impact assessment of article 35. ## Three things not to do **Biometric surveillance during the exam.** Facial recognition to confirm that the same person is still there, gaze tracking, emotion detection, keystroke dynamics. Automated analysis of a candidate’s behaviour carries a high risk of false positives, and it turns the exam into an exercise in looking “normal” for the tool. When it works on the body, it processes special-category data under article 9 of the GDPR, and it is hard to argue as necessary under article 5. Nothing on the list above needs it. **Automated verdicts.** A risk score, a count of exits, an alert are inputs for a person. A system that closes an exam or fails a candidate on its own takes an automated decision under article 22 of the GDPR, which restricts it precisely because the effect on the candidate is significant. The examiner decides, and answers for the decision. **Filming beyond need.** A 360° scan of the room, a continuous screen video when event-triggered captures would do, a second camera pointed at the desk, recordings kept for a year “just in case”. Each of these costs every honest candidate something, and each has to be justified in the impact assessment of article 35, which looks at the set-up as a whole, not measure by measure. ## Where Evalmee stands Measures 3 to 12 exist in Evalmee, in a standard browser, with an optional desktop app and no lockdown browser. The paper is never downloadable in advance. Full screen is mandatory, screen exits are counted with their cumulative duration, copy and paste and a second screen are detected, and a screenshot is taken on four events only: leaving the window, switching to another application, leaving full screen, a suspicious browser extension. Identity is checked by the examiner, who compares the photos of the document and the face; the standard product has no facial recognition, and the check requires video proctoring, which the institution enables, never by default. Webcam and microphone recordings are deleted after three months by default; screenshots follow the retention period of submissions, set by the examining institution. From those events the platform computes a [fraud](/en/assessment-glossary/#exam-misconduct) risk score, and the score is not a verdict: it closes no exam and reports nobody; the examiner reads the timestamped events and decides. The detail is on the [exam monitoring page](/en/product/monitor/), the commitments on hosting, encryption and retention on the [security page](/en/security/). What Evalmee does not see is worth stating just as plainly: a phone on the desk, a person out of frame, a second device. No online tool does. ## No measure makes cheating impossible The honest answer to the question in the title is that a few measures, combined, make cheating costly, uncertain and traceable, and that this is what an exam needs. The choice follows the stakes. A mock exam or a formative test needs measures 1 to 6, which collect no personal data. A graded exam or a certification adds 7 to 11. Measure 12 is for exams whose stakes justify it and where an in-person session is not an option. The most defensible set-up is not the most complete one; it is the one you can still explain to the candidate who asks why the camera was on. ## Frequently asked questions ### Can cheating in online exams be prevented completely? No. Every measure in this article closes one route and leaves others open: a phone on the desk, a person out of frame, an earpiece. A few measures combined make cheating costly, uncertain and traceable, which is what an exam needs. They are chosen according to the stakes of the exam, not stacked by default. ### Is webcam proctoring legal under the GDPR? Yes, under conditions. Real-time video and audio monitoring can be proportionate for a high-stakes exam, provided candidates are informed in advance (article 13 of the GDPR), the retention period is set before the system goes live, access to the recordings is restricted, no automated analysis of the candidate’s behaviour is performed, and the impact assessment of article 35 has been run beforehand. A mock exam should run without it. ### Does a lockdown browser stop cheating? It blocks the other applications on the machine it runs on. It sees nothing of a phone, a second device or a person in the room, and it excludes candidates whose computer cannot install it. Detecting screen exits, copy and paste and a second screen in a standard browser covers much of the same ground without the installation. ### Which measures are enough for a low-stakes test? Clear rules, questions that cannot be looked up, a paper that is not downloadable in advance, a random draw with a shuffled order, and a time limit. None of them collects personal data. Proctoring measures are for exams whose stakes justify them. ### Can a fraud risk score be used to fail a candidate? Not on its own. A score or an alert is information for the examiner, who reads the events behind it and decides. A decision taken by a machine alone, with a legal or similarly significant effect on the candidate, is what article 22 of the GDPR restricts.