AI watches. Humans decide.
AI proctoring at scale is easy to do badly — false positives, black-box penalties, auto-rejections. 5Profiler’s AI layer detects and documents; the decision always belongs to a human, and the candidate can always contest it.
Flags human-reviewed
100%
By design
0 auto-rejections
Proportionate stakes
4 tiers
Engineered against false positives
The worst proctoring outcome isn’t a missed cheat — it’s a good candidate wrongly flagged. The AI layer is grounded, escalated carefully, and never the final word.
- Vision with a second opinion — uncertain detections escalate to a stronger model before they ever reach a reviewer.
- Cost- and privacy-bounded — per-candidate caps on AI checks; names and addresses redacted before any prompt leaves the platform.
- Accommodations-aware — religious dress and accessibility declarations adjust detection automatically — dignity by default.
- Contestable — candidates see their own integrity result and can formally appeal it.
Detect → evidence → score → human
1 · Detect
Face presence, gaze, multiple faces, second voices, device signals — sampled intelligently, not filmed wall-to-wall.
2 · Attach evidence
Every detection ships its clip or frame — reviewers see what happened, not that “something” did.
3 · Score transparently
Open penalty math into one integrity score — your team can tune the weights per campaign.
4 · Review humanly
A triage queue ranks what deserves attention; borderline calls land with people, never a threshold.
The ceiling fan incident
A vision model once read a ceiling-mounted object as a phone in hand — a 24-point penalty from a photograph of a fan. That class of failure is why the pipeline exists: grounding checks, second-opinion escalation, transparent penalties and human review mean one bad frame can’t cost a candidate a job.
Questions we always get
How do you handle false positives?
Detector health is tracked org-wide, uncertain calls escalate to a second model, penalties are transparent and tunable, and humans review the edges — with candidate contest as the final safety net.
Is candidate data used to train AI?
No. Vision checks are inference-only, PII is redacted at the prompt boundary, and evidence lives under your retention policy.
Does AI ever reject a candidate?
Never. AI detects and documents; people decide — and that posture is written on the report.
Go deeper in the research library: integrity flags and false positives · AI-resistant assessment design.
People also evaluate
AI-assisted. Human-decided.
A 30-minute walkthrough with your roles, not a canned deck — and a sample report to keep.