Picture the job honestly. A proctor is watching a grid of small video tiles — sometimes a dozen at once. Each tile is a webcam pointed at a face, a slice of a room, and not much else. The proctor has to notice the one glance that lasts a beat too long, across twelve people, for two hours, without blinking past the moment it happens.

People are bad at this. Not because proctors are careless — because the task is close to impossible, and the feed they are given was thin to begin with.

What a webcam frame contains

A standard proctoring tile shows a face and whatever is directly behind it. It does not show the desk surface, the lap, the wall the candidate is facing, or the phone lying just out of shot. The most common cheating aids live precisely in those blind spots, because anyone setting out to cheat puts them there on purpose.

So the proctor is asked to infer wrongdoing from the one thing the camera does capture: behaviour. Eyes drifting. A pause. A lean. And behaviour is a famously noisy signal.

The false-positive tax

Anxious people look like they are hiding something. Neurodivergent candidates often do not hold a webcam's idea of "normal" eye contact. People think with their eyes up and to the side. Flag all of that and you generate a flood of accusations against honest test-takers — and every wrong accusation is expensive, not just unfair. Where programmes bolt automated face or gaze analysis on top of the human, the error is not even evenly distributed: NIST's testing has repeatedly found demographic differences in face-analysis accuracy.

An integrity system that cannot tell anxiety from cheating does not protect the exam. It just redistributes suspicion onto the people least able to push back.

This is the quiet cost of leaning on behavioural judgement: it taxes the honest majority to occasionally catch a careless cheat, while the careful ones stay perfectly still and look exactly like everyone else.

The part that is structurally invisible

Here is the harder truth. Even a perfect proctor, fully attentive, cannot see what was removed from the feed before it reached them. A screen overlay that excludes itself from capture is not a thing the proctor missed — it is a thing that was never in the image. We pulled that apart in why remote proctoring is blind, and it is the difference between "watch more carefully" and "you are watching the wrong picture."

No amount of attention recovers information that is not there. That is not a training problem. It is an architecture problem.

Where this leaves the human

None of this means proctors are useless. A present, capable human is excellent at the things humans are good at — judgement on edge cases, de-escalation, deciding when something genuinely warrants a second look. The mistake is asking them to be a sensor instead of a decision-maker.

The better division of labour: let infrastructure capture a complete, honest picture — the goal behind OroLink — and let analytics surface patterns across sessions, as in the fingerprint that exam fraud leaves behind. Then hand the human the few cases that actually need a person. You get fewer false accusations and catch more of the organised fraud that a tile-by-tile watch was never going to see.