How the AI proctoring risk score works.

What the score means

The score is a weighted sum of independent proctoring signals. Each signal contributes points up to a cap. The total is bounded at 100 and bucketed into three bands the recruiter UI surfaces as colour-coded badges:

• Low (0–29)

  No signal or only minor signals.

• Medium (30–69)

  Worth a manual review of the breakdown.

• High (70–100)

  Multiple strong signals fired — investigate before deciding.

Signals and weights

The current heuristic engine version is heuristic-v1. Each contribution shows up in the recruiter breakdown with the raw event timestamps so the score is fully traceable, never a black box.

What the score is not

• Not a hiring decision. The score is advisory input for the recruiter. Final pass/fail and hiring outcomes are made by humans, not by the score.
• Not a profiling decision under GDPR Article 22. Test Utopia does not produce automated decisions with legal or similarly significant effects. The score is shown to the recruiter as an aid; they choose how to use it.
• Not stored beyond the attempt. The risk score and breakdown live on the same test_attempts row as the rest of the candidate’s data and are deleted on the same retention schedule (or sooner on a Data Subject Request — see Privacy).
• Not a behavioural model trained on candidate data. The current engine is deterministic and rule-based; the same inputs always produce the same score. No machine-learning model is trained on candidate data without an explicit, future, separately-disclosed methodology change.
• Not enabled by default. Customers must opt-in per test (Test settings → AI Proctoring Risk Score). Tests without the toggle do not produce a score; existing attempts that completed before the feature shipped do not retroactively get one.

Candidate notice

When a customer enables AI proctoring on a test, the proctoring telemetry is the same telemetry the existing test-protection toggles produce — fullscreen, tab switching, paste detection, geolocation, and so on. The risk score is a derived view of that telemetry, not new data collection. Candidates see the protection features that are enabled before they begin the test, in the test introduction screen.

Roadmap

• Phase 1 — heuristic-v1 (shipped). The signals and weights described on this page.
• Phase 2 — LLM code-authenticity scoring (shipped). On every completed attempt of a proctoring-enabled test, the candidate’s submission history is sent to a large language model provider with the question prompt, starter code, every code submission attempt, and the timestamps of any paste events captured during the attempt. The model returns a 0–100 authenticity score plus a short reasoning paragraph; we combine it with the Phase 1 heuristic via final = max(heuristic, llm) so neither path alone can bury a strong signal. The LLM call is async (runs in a separate worker process), never blocks attempt completion, and is capped by a per-tenant daily budget. What the call sees: final code, every submission’s code text, paste-event timestamps. What the call does not see: keystroke-level telemetry, candidate identity, webcam, microphone.