How to conduct online exams securely: preventing cheating in 2026

What does it mean to conduct an online exam securely?

It means that after the exam you can answer three questions with evidence: who took it, what they had access to, and what happened during the session. Security here is not the absence of cheating. It is the ability to detect it, bound its impact and prove your case if a result is disputed.

The failure mode of most online exams is not a dramatic hack. It is a result that cannot be defended: a candidate contests a score, the institution has a screenshot and a hunch, and the outcome turns on who argues better. Every control below exists to replace a hunch with a record.

What is remote proctoring, and what types are there?

Remote proctoring is the supervision of a candidate taking an exam outside a physical exam hall, using their webcam, microphone, screen and activity signals. There are four models in common use, and they trade cost against certainty in opposite directions.

How does AI proctoring detect cheating?

It does not “detect cheating”. It detects deviations from the expected pattern of an honest session and attaches a confidence score to each. Understanding the difference is what keeps an integrity process defensible.

What is a secure or lockdown browser, and do you need one?

A lockdown browser is a restricted application that runs the exam full-screen and blocks the things a normal browser allows: other tabs, screen capture, virtual machines, copy-paste, screen sharing and commonly, external displays. It shifts you from detecting that a candidate opened another window to preventing it.

You need one when the exam is high-stakes and the answers are easily searchable. You can often skip it for open-book or long-form exams, where a quick search is worth little and forcing an install on every candidate's device costs more in support tickets than it buys in security. Browser-based monitoring with focus detection is the middle ground.

How do you verify that the right person is taking the exam?

Identity is checked at three moments, not one. Checking only at the start is the most common gap in online exam design. It leaves the whole session unverified after the first minute.

1. Before entry: unique per-candidate credentials or links (never a shared password), photo-ID capture, and a live selfie matched against the ID.
2. During the session: periodic presence checks, face continuity, single-face verification, and session binding so the same login cannot be active on two devices.
3. In the answers themselves: at least one response type that carries the candidate's identity intrinsically, such as a spoken or recorded answer. This is the check that survives every form of screen-level trickery.

Does randomisation actually reduce cheating?

Yes, and it is the cheapest control on this list. Randomisation changes the economics: it converts a leak that would compromise everyone into one that compromises a single candidate.

• Item pools. Draw each candidate's paper at run time from a bank several times larger than the paper. A 60-item exam drawn from 300 calibrated items means no two candidates see the same set.
• Option shuffling. Randomise answer-option order so “the answer is C” is meaningless to a neighbour.
• Section and question order. Shuffle at both levels to break screen-sharing coordination.
• Sectional timers and one-way navigation. Removes the window in which a candidate photographs the whole paper and sends it out for answers.
• Staggered start times where the calendar allows, so the cohort is never all sitting the same items simultaneously.

The prerequisite is an item bank large enough and calibrated well enough that different draws are comparably difficult. Randomising a pool of wildly uneven items trades integrity for unfairness, which is not a good trade.

How do you design questions that are hard to cheat on?

The most durable defence is not surveillance. It is asking for something a search engine, a chatbot or a person in the next room cannot supply on demand.

How do you conduct a secure online exam, step by step?

A checklist you can run for any exam, from a 30-student quiz to a 5,000-candidate certification.

1. Classify the stakes. Decide up front whether this is low-stakes (formative), medium, or high-stakes. Every control below is scaled by this answer. Do not proctor a practice quiz.
2. Build an item bank, not a paper. Author 3–5× the number of items the exam needs, tagged by topic and difficulty.
3. Mix response formats. Include at least one item type that carries authorship evidence: audio, video, or worked long-form.
4. Configure randomisation. Pool draw, option shuffle, section order, sectional timers.
5. Set the access model. Unique per-candidate credentials or links, exam window, single active session per candidate, and no shared passwords.
6. Choose the proctoring level and decide, in writing, which signals you will act on and at what confidence threshold.
7. Get consent and publish the rules. Tell candidates exactly what is recorded, for how long it is retained, who reviews it and how to appeal. Say it before they book, not on the exam screen.
8. Run a mandatory system check. Camera, microphone, bandwidth and browser, at least 24 hours before the exam, so technical failures do not become integrity flags.
9. Do a live dry run with a handful of real users on real devices, including one low-bandwidth connection and one older phone.
10. Monitor during the exam and staff a support channel. Most mid-exam incidents are technical, and an unanswered candidate is a future dispute.
11. Review flags with a human after the exam. Record the decision, the evidence and the reason for each flagged session.
12. Retain the audit trail (event log, recordings, response timings) for the length of your appeals window, then delete it on schedule.

How do you stay fair, legal and defensible?

Consent and privacy

Proctoring processes biometric and household-environment data. Obtain explicit, informed consent in advance; state the retention period and stick to it; restrict reviewer access; and offer an alternative arrangement where a candidate genuinely cannot be recorded. XAM publishes a standard proctoring consent notice you can adapt as a starting point.

Accessibility and equity

Extra time, screen-reader compatibility and keyboard navigation must work with your lockdown settings, not against them. Assistive technology is a frequent source of false flags. Assume some candidates are on shared devices and unstable connections, and design the connectivity-loss recovery path deliberately.

Due process

Publish the appeal route, give the candidate the specific evidence against them, and let a person who was not the original reviewer decide the appeal. An integrity process without due process will not survive its first serious challenge.

XAM by Ankor, an assessment platform, implements these layers in one engine: authoring and item banks, randomised delivery, AI proctoring with confidence-scored flags and session replay, audio and video answers as a native question type, auto-evaluation and branded reports. The free tier runs 20 real sessions with no card, which is enough to test the checklist above end to end. See pricing or talk to us about an institutional rollout.