Why AI is necessary for online exam security
Manual video review of thousands of concurrent remote exams is impractical. A university running a semester exam for 10,000 students over two hours would require hundreds of human proctors watching simultaneously.
AI makes large-scale proctoring economically viable by automatically analysing every session in parallel and escalating only the flagged moments for human review.
Computer vision: what the camera sees
Face presence and recognition: The system confirms the registered candidate is visible throughout the session. If the face leaves the frame for more than a configurable threshold, the session is flagged.
Multiple face detection: A second person appearing in the webcam frame (someone reading answers aloud to the candidate, for example) is one of the clearest fraud signals and is reliably caught by object detection models.
Eye-gaze estimation: Head pose and iris tracking infer whether the candidate is looking at the screen, looking down at notes, or repeatedly glancing off-camera.
Object detection: Phones, books, paper notes, and second monitors can be detected in the camera frame using trained object detection models.
Audio analysis: what the microphone hears
Voice activity detection identifies when a second voice is present in the room, which may indicate collaboration or external coaching.
Keyword detection: Some advanced systems listen for phrases common in exam misconduct, though this is less commonly deployed due to accuracy and privacy concerns.
Background noise profiling: Sudden changes in ambient noise — a phone notification, a knock — are logged as events.
Browser and screen monitoring
Tab and window focus events: Any time the candidate navigates away from the exam browser tab, the event is recorded with a timestamp.
Clipboard monitoring: Copy-paste attempts are logged, which can indicate the candidate is moving content out of or into the exam interface.
URL and application access: Lockdown browser modes prevent navigation entirely; standard modes log the URLs visited.
Screen-sharing detection: The system detects if the candidate has initiated screen-sharing, which could indicate they are sharing the exam with an accomplice remotely.
Behavioural anomaly scoring
Individual signals can be ambiguous. Looking off-camera to think is normal. Looking away every 30 seconds may not be.
Modern proctoring systems aggregate signals into an anomaly score for the full session, weighting events by frequency, duration, and combination.
Proctyx generates a per-candidate integrity report showing flagged events on a timeline, so reviewers can watch the relevant clips rather than reviewing an entire recording.
Reducing false positives
False positives — flagging honest candidates — are the most significant risk in AI proctoring. Poor lighting, glasses reflection, and natural thinking behaviour (looking up or to the side) can all trigger face-absence flags.
Best practices: run a system check before the exam so candidates can verify camera and lighting; set thresholds conservatively; never auto-disqualify based on AI flags alone; use human review as the final step.
Proctyx allows administrators to configure sensitivity levels per exam type — low-stakes practice tests can use minimal monitoring while high-stakes certification exams use full AI analysis with human review.
FAQ
Can AI proctoring be fooled?
Determined candidates can attempt countermeasures (phone cameras, virtual machines), but most common cheating methods are detectable. Combining AI monitoring with question randomisation and time limits makes cheating significantly harder.
Does AI proctoring flag innocent behaviour?
It can. That is why human review is essential. AI generates flags; trained reviewers decide whether the evidence constitutes misconduct.
What happens when cheating is detected?
The proctoring platform flags the event and notifies the institution. A human reviewer examines the evidence and the institution applies its own academic integrity policy.
Want to see Proctyx in action?