The useful question about AI proctoring is what a signal can support. A missing-face alert says the system did not detect a face under the conditions of that check. It does not establish that the student left to get help. This guide explains the current browser exam implementation and gives a short rehearsal procedure.
The October 2026 revision was prepared from the exam source and linked library documentation. It is not an accuracy benchmark, an independent security audit or evidence from a classroom trial.
From camera input to an event log
- The browser requests media permission when monitoring is enabled. The institution can disable AI proctoring for a session.
- JavaScript models process camera frames on the candidate’s device. The exam uses face-api.js and a TensorFlow.js object-detection model.
- Checks apply thresholds, repetition rules and cooldowns before recording events. This reduces some transient alerts, but does not guarantee that recorded events are correct.
- Counts and timestamped events are written to the session record for the proctor. Network problems may affect when those updates arrive.
Face detection is different from identity verification
The face-api.js project provides distinct detection, landmark and recognition capabilities. Using a face detector does not itself match a student to an identity document. In ZYNQO, face and landmark checks support presence and related monitoring signals; an institution still needs its own enrolment and identity process.
A student leaning to read, camera glare, movement or an unsuitable camera position can interfere with a check. A second face may be a household interruption. A gaze-related signal cannot establish what someone was looking at or whether they used prohibited material.
Object scores are not misconduct probabilities
COCO-SSD returns object classes, scores and bounding boxes. The exam checks selected classes, including phones and some other devices or materials. The current code requires repeated qualifying detections for its device alert. A permitted book or an incorrectly classified object can still require review.
A score above a configured threshold is a model output, not an independently measured accuracy percentage. There is no published ZYNQO classroom benchmark here that establishes a false-positive rate. Objects outside the camera view may be missed entirely.
Local processing and stored information
The exam code processes webcam frames and, where microphone access is granted, sound-level signals in the browser. It does not include an audio/video recording-upload flow. The session still sends answers, identifiers, timestamps and monitoring events to Firebase. Local media processing does not mean that no personal information leaves the device.
Permissions, account state and the exam setup affect the experience. Read the privacy policy, and ask the institution who can access the event record and when it will be deleted. Do not assume that choosing a browser-based tool alone establishes legal compliance.
Alerts can trigger a configured submission limit
The current exam supports an optional violation limit. Reaching that configured limit can submit the exam automatically. Institutions should check that setting before an exam, explain it to candidates, and plan how to review incorrect alerts. A submitted attempt is not proof of misconduct.
A repeatable practice session
Use a dummy exam and a consenting tester on the actual device and browser intended for the sitting. Do not test disruptive controls during a graded exam.
- Check camera framing and ordinary reading, typing and permitted assistive technology. Note alerts during normal work.
- Try a brief, permitted movement and return. Compare the candidate’s experience with the event record.
- Place a permitted calculator or book in the working position. If an alert appears, agree a suitable configuration or alternative arrangement.
- Test the automatic-submission setting, if enabled, and confirm the result appears for the proctor.
- Record browser, device, lighting, settings and outcome in the readiness worksheet.
One successful rehearsal establishes only what happened in that rehearsal. Changes to the device, browser, models, lighting or configuration can change the result. The practical outcome is a documented decision about whether this setup is suitable, not a claim that the system is infallible.