An Indian university moved its semester assessments online during the pandemic, kept them online because the logistics were cheaper, and then spent three years discovering that a flagged proctoring report is not evidence. Every term the examination committee sat with a spreadsheet of flags, watched video clips, and argued about whether a student looking down for eleven seconds was cheating or reading the question on paper.
That experience is now common enough that the sector has moved on from asking whether proctoring catches cheating to asking what a defensible remote assessment looks like.
The Limits of Watching a Camera
Automated proctoring produces signals, not verdicts. Gaze deviation, second face detection, tab switching, audio events, and typing cadence all correlate weakly with misconduct and strongly with ordinary life. A student in a shared room in a tier-two city triggers audio flags all evening because there is a family in the next room.
Published false positive rates for automated flagging systems commonly sit in the range of 10 to 30 percent of sessions depending on how sensitivity is tuned, which means a 2,000-student exam can generate several hundred items for human review. Most institutions do not staff for that, so flags get skimmed or ignored, and the deterrent value collapses along with the review process.
There is also a fairness problem that has been documented repeatedly. Gaze and face detection perform unevenly across skin tones, and students with certain disabilities or neurodivergent movement patterns get flagged at higher rates. If your appeals process cannot answer that, your credential is exposed.
Design the Assessment So Cheating Is Expensive
The more durable approach is to change what is being asked.
- Item banks with randomised selection. Serving 30 items drawn from a pool of 300 by topic weight means no two students see the same paper, and a leaked screenshot is worth very little.
- Parameterised numerical questions. Generate the numbers per student. The method is the answer, and a shared final value is useless.
- Open-book, application-level items. If the answer is retrievable in ten seconds, the question was testing retrieval, not competence. Case analysis and applied reasoning are far harder to outsource.
- Time pressure calibrated to the task. Enough time to think, not enough to consult a third party mid-question. Per-item timers work better than a single paper-level clock.
- Stakes distributed across the term. A single 100-mark endpoint invites desperation. Continuous assessment reduces the payoff of any one attempt.
- Identity verified once, properly. A checked government ID at session start is worth more than continuous face matching that fails when the light changes.
The Governance Layer Nobody Budgets For
Whatever the technology, three things need to exist before the first online exam runs. There has to be a written policy that tells students exactly what is monitored, what is retained, and for how long. Under India’s Digital Personal Data Protection Act, biometric and video capture of students, many of whom are minors in the school segment, requires a lawful basis and a retention limit you can actually defend. Ninety days of raw session video sitting in an unmanaged bucket is a liability, not a control.
There has to be a human review path with a named owner and a service level, because a flag that nobody reviews within the results timeline is not part of the process. And there has to be an appeal route where a student can see what was recorded about them and respond to it. Institutions that skip this end up litigating individual cases rather than defending a system.
At Invexa, we treat assessment integrity as an examination design problem with a software component, not the other way round, because the systems that hold up under challenge are the ones where the paper itself was hard to game.