AI Proctoring Analysis

AI Detections automatically reviews the webcam recordings of a completed test session and flags moments that deserve a human look: a second person in the room, a different face than the ID photo, the candidate looking away or leaving, a phone or book in view, talking, or a muted microphone. Every finding is pinned to the exact photo or video second, so a reviewer verifies in seconds what would take hours of scrubbing. The AI never fails a candidate by itself — it collects evidence, computes a risk level, and leaves every decision to you.

How it works

1. During the exam, the candidate's browser records webcam photos (every 20 seconds) or video (continuous, with audio), depending on your step's security settings.
2. When the session completes, analysis starts automatically (if AI Detections is enabled on the step). Video is sampled at your configured interval (default: one frame every 10 seconds); every sampled frame runs the checks you enabled.
3. When something is flagged on video, the AI double-checks it: it re-examines the surrounding seconds more densely, discards one-frame glitches (a blink, a compression artifact), and records how long the incident actually lasted. You can turn this verification off for faster, rougher results — not recommended.
4. Findings appear as detections on the session, each with its evidence (what was seen, how confident, when, for how long). The session gets a traffic-light verdict: 🔴 likely cheating, 🟡 suspicious, 🟢 clean.
5. A reviewer opens the session, watches the flagged moments, approves or declines each finding, optionally adds manual findings, and finalizes the session with their own verdict. The AI's verdict is a starting point, never the final word.

What each detection means

• Person count — Nobody in front of the camera, or more than one person. Typical innocent causes: bathroom break; family member passing behind.
• Identity check — The face doesn't match the ID photo taken at the start. Typical innocent causes: poor ID photo quality, dramatic lighting change.
• Face position — Head turned away or eyes off-screen (looking at notes, a second screen, a phone). Typical innocent causes: thinking while looking at the ceiling; second monitor showing the exam itself.
• Forbidden objects — A phone or book visible in frame. Typical innocent causes: phone used as an allowed calculator; decorative bookshelf.
• Voice activity — Someone talking during the exam. Typical innocent causes: reading questions aloud; background TV.
• Muted microphone — The mic recorded nothing at all. Typical innocent causes: hardware or permission problem rather than intent.

Video sessions support all six detections; photo sessions support the first four (no audio). Screen recordings are not AI-analyzed — reviewers flag those manually. Person count and identity findings are treated as strong (red) evidence; the rest as moderate (yellow). A face-position finding escalates to red when the deviation is extreme (default: 70°+ or hard gaze) and verified as sustained (10 seconds or longer) — a prolonged stare at a second screen reads very differently from a momentary glance.

The session verdict

The session's traffic light comes from a risk score, not from any single finding:

• Every non-declined finding contributes points — stronger kinds contribute more, human-approved findings count 1.5×, and longer incidents count up to 2×.
• Repeats of the same signal saturate: 40 "looked away" findings caused by bad lighting won't turn a session red by themselves. Different signals compound: a few glances plus a phone plus talking will.
• The score is normalized per monitored hour, so long exams aren't penalized for accumulating noise.
• 🔴 red = high score, or a proctor-confirmed strong finding (approved "different person" / "second person", or a manual cheating flag — these are red regardless of score). 🟡 yellow = moderate score, or any unreviewed strong finding (strong evidence never silently disappears into green). 🟢 green = nothing meaningful.
• Reviewing findings moves the light: declining false positives removes their points (and they will never come back on future runs); approving real ones increases their weight. Your manual flag always wins — switch the session to manual mode and set your own verdict at any time; the AI never overwrites it.

Usage scenarios

• High-stakes certification: Video + all detections + defaults; review every yellow/red; approve/decline each finding; finalize with manual flag. The audit history (who reviewed what, when) is your defensible record.
• High-volume hiring screen: Photo mode + person count + identity check only (cheap, fast); sort the session list by verdict; only open reds.
• Open-book or calculator-allowed exams: Disable forbidden objects (or raise its confidence), raise the face-position angle — candidates legitimately look down at materials.
• Noisy environments (home candidates): Disable voice activity or accept and bulk-decline; keep muted-mic on to catch disabled microphones.
• Investigating a complaint: Re-run with a denser interval and stricter thresholds on the specific session; new findings merge with the existing review instead of replacing it.
• Budget-conscious: Run exams without AI; retroactively enable it only on the sessions where the score matters (top candidates, disputed results).

Good to know

• Does the AI fail candidates automatically? No. It flags and scores; humans decide. Unreviewed strong findings keep a session at least yellow so nothing slips through unseen.
• When does analysis run? Automatically when the session completes (if enabled), typically finishing within minutes; long video sessions take longer. You can watch progress live.
• What can't it see? Screen content (use screen recordings + manual review), a second device outside the frame, and anything before the webcam starts.
• Photo vs video? Photo is lighter on candidates' bandwidth and cheaper; video adds audio checks, incident durations, and red-escalation of sustained behavior. High-stakes → video.
• Who can do what? Admins, managers, and operators run and enable analyses. Task-scoped users can re-run analyses on tasks they can edit. Enabling retroactively (which costs credits) is admin/manager/operator only.
• Changed your detection settings? New settings apply to future runs. Re-run the analysis to apply them to an existing session — pending findings from disabled kinds are cleaned up automatically; reviewed ones stay.
• Can a declined finding come back? No — declines are permanent per finding, survive every re-run, and keep their audit trail.