Every proctoring product raises flags. The software notices a face leaving the frame, a voice in the room, a second display, a tool it recognizes. That part of the job is well understood. Our guide to AI proctoring covers what the models look for. This post covers the part people discuss less: what should happen in the minutes after a flag fires, and who is accountable for it.
Our position is simple. An AI flag is never a finding. A trained person looks at it in context, records a decision with a reason, and the institution makes the final call.
A flag is a guess about where to look
Think about the ordinary things that happen during a two-hour exam at home. A candidate looks up at the ceiling while working through a problem. A younger sibling opens the door. The light changes as clouds pass. A student who reads questions aloud to themselves does exactly that. Someone with a tremor moves in ways a model was never trained on. Each of these can raise a flag, and none of them is misconduct.
The software cannot tell these apart from genuine concerns, because it does not know the candidate, the room or the accommodation the student was granted. A person watching the clip at the flagged moment usually can, and that is why review matters.
Who pays when flags go straight to penalties
When automated flags lead directly to a zero, an extra viva or a referral, the cost of every false alarm lands on an honest student. The exam office pays as well, because it then has to defend a decision nobody actually made. Appeal panels ask three questions: what happened, what was seen, and who decided. “The system flagged it” answers none of them.
There is a quieter cost too. Faculty who see a long list of unexplained alerts learn to ignore them, which means the real incidents get the same shrug as the false ones.
What a human decision should look like
Saying “a person reviews it” is easy. The review itself needs a clear shape, and we think it has four parts:
- Context. The reviewer sees webcam, screen and events on one timeline, opened at the flagged moment, with what happened before and after.
- A required reason. No decision is recorded without one. Reasons expose inconsistent calls and give the candidate something concrete to respond to.
- Separation of roles. The reviewer records a disposition. The institution, through its own committee, makes the finding. Nobody imposes a penalty before that recorded decision.
- A record that survives. Every view, decision and change is logged in a tamper-evident audit trail, and the case exports as one evidence pack.
The mechanics of this on Proctorly, including the reviewer console and the review priority score, are set out on our human review proctoring page.
Being honest about the limits of the recording
A human decision is only as fair as the evidence in front of it. A camera and a screen recording cannot see every way of getting help, and a reviewer can only judge what was recorded. When a camera fails or an upload drops, that is a technical failure, and Proctorly never treats a technical failure as misconduct. We also publish what each setup can and cannot observe on what we can and cannot see, so reviewers and committees can weigh evidence against what was actually observable.
What we do not ask the AI to do
We draw some lines explicitly. The software on Proctorly does not score emotion, personality or “suspicion”. It does not infer intent from a face or a voice. The review priority score helps reviewers decide which cases to open first, and it is labeled a priority for review, never a finding. Unusual audio and movement are raised as signals only.
A flag exists so that a fair, recorded review can happen. Everything after it is a human responsibility.
What this asks of institutions
Human review takes work. Someone has to do it quickly, during the busiest weeks of the year. Some institutions review with their own trained staff at the Recorded level. Others ask Proctorly reviewers to take the first look, working 24×7 during exams in English, with one reviewer for every 40 candidates. Either way, the institution names who decides and how candidates appeal. We think that is the right division of labor, and it is the one most likely to hold up when a result is challenged.
Frequently asked questions
Does human review slow down results?
At the Every flag reviewed level, Proctorly reviewers review 95 of every 100 flags within 5 minutes. The rest are still reviewed, just later, and institutions can set their own turnaround.
Can an automated flag ever fail a candidate on Proctorly?
No. No automated flag fails a candidate or imposes a sanction. Confirmed incidents go to a recorded committee decision, with no penalty before it.
See a flag go from alert to decision in a full walkthrough. Watch how it works.