AI proctoring has come a long way from the early days of rigid webcam monitoring that flagged any movement as suspicious. Today's systems are built with a core principle: the 98% of students who play fair should never feel like suspects.
The false positive problem
Early proctoring systems had false positive rates as high as 12–18%. A candidate scratching their nose or glancing at a sticky note on their monitor would trigger a flag. The result? Mass appeals, disillusioned institutions, and eroded trust in AI proctoring as a whole.
How on-device ML changed the game
Moving inference to the candidate's device (rather than streaming video to cloud servers) enabled faster, more contextual decisions. Instead of flagging a single off-screen glance, the model now looks at patterns: how many times, for how long, and in what sequence.
Configurable thresholds
ExamRankers lets institutions set minimum durations before a flag is raised. The default: flag a gaze-away event only if it lasts more than 5 seconds and occurs more than 3 times. This alone eliminates over 70% of false positives compared to threshold-free systems.
Results from the field
After piloting with 500 students at EduVision Academy, the false positive rate was 1.8%, all in the Low severity category. Not one candidate filed a formal objection. The institution subsequently rolled out to 50,000 students.