Proctoring flags are not findings: how to review automated alerts without convicting the innocent

Start with the base rate, every time

If two percent of a cohort cheats and your flags touch twenty percent of sittings, simple arithmetic says most flagged students did nothing wrong, however good the model feels. Reviewers need that number in front of them, per exam, per flag type. Gaze flags fire constantly and corroborate little. A second voice discussing the questions is rare and means more. Weight reviewer attention accordingly instead of giving every flag type the same red badge and the same fear.

Demand your vendor's validation data by flag type and by student group, and log your own outcomes: flags raised, flags cleared on review, cases upheld on appeal. A vendor who cannot tell you a flag's false positive rate in conditions like yours is selling you the liability and keeping the certainty.

A review protocol humans can actually follow

Two reviewers for any case that could become a finding, working from the clip plus thirty seconds either side, the exam's own rules on breaks, materials, and assistive technology, and the student's prior flags, or better, the absence of them. Reviewers record what the clip shows in factual language, never what it proves. Escalation needs a behaviour that matches a defined offence: consulting a phone, a second person supplying answers, a forbidden resource open. Anything less clears, and the clearance is recorded with the same care as an escalation.

Assistive technology and access arrangements must be visible to reviewers inside the review tool, not in a separate system they have to remember to check. A screen reader flagged as an unauthorised application, reviewed by someone who cannot see the approved adjustment, is a process failure your institution chose when it bought the workflow.

Appeals are an audit of your thresholds

Track appeal rates and overturn rates by flag type, by course, and by student characteristics your equity monitoring already collects. An overturn rate concentrated in one group means your threshold is not neutral in practice, whatever the policy says. Publish the headline numbers to the committee that owns the policy. Systems that are never audited drift toward severity because severity feels safe to administrators and is paid for by students.

Give students the clip and the specific allegation when a case proceeds. A student who cannot see the evidence cannot answer it, and a process that hides its evidence will lose the appeals that matter and deserve to.

Clusters of flags at the same question often mean the question is ambiguous or the permitted materials rule is unclear, not that a cohort conspired at minute forty. Review flag clusters as exam feedback first. The cheapest integrity improvement in most courses is a clearer paper, a stated calculator and notes rule, and a practice sitting where students learn what the software sees before the stakes are real.

Use proctoring data to fix exams, not only to police them

Frequently asked questions