Anti-cheat Scoring

Catch coached, AI-assisted, and off-camera candidates.

2025 made cheating accessible to anyone with a second monitor and ChatGPT. Raffi's anti-cheat layer is built into every interview — risk-scored 0–100 with the specific signals that triggered it.

Signals Raffi tracks per interview

  • Response latency patterns (long pause → fast burst = AI assist)
  • Linguistic fingerprinting (LLM-style sentence structures)
  • CV cross-reference (claims that don't match the resume)
  • Eye-movement and gaze direction (off-camera prompts)
  • Audio anomalies (background voices, headphone tells)
  • Retake fingerprinting (matches across retake attempts)
What it does

The capability, broken down.

Real-time scoring during the interview

Anti-cheat doesn't wait until the call ends. Raffi scores signals live and adapts its questions — if it detects a pattern that looks coached, it asks a follow-up the candidate couldn't have prepared for. Real talent answers fine; coaching breaks.

Post-interview audit

After the call ends, Raffi runs a full audit: linguistic patterns vs known LLM outputs, fact-checks against the CV and LinkedIn, retake fingerprinting if the candidate redid sections. You see the final risk score with every transcript.

Flags are explainable, not opaque

When the risk score is high, you see exactly why — 'unusually long pause at 03:42 followed by perfect technical answer' or 'claimed 5 years at Acme but LinkedIn shows 18 months'. You make the call, not Raffi.

Calibrated against real interviews

Anti-cheat thresholds are tuned against 100,000+ real interviews. False positives are visible to you — you can override, and the model learns from your overrides. The goal: catch the 5% who would have wasted your time, not flag the 95% who are real.

Under the hood

Technical specifics.

Risk score range0–100 (0 = clean, 100 = high risk)
Signal categories6 (latency, linguistic, CV, gaze, audio, retake)
Real-time vs post-auditBoth — real-time adapts questions, post-audit finalizes score
Override flowYes — recruiter can mark false positive, model learns
TransparencyEvery flag is explained with the specific evidence
Training set100,000+ interviews, continuously updated
FAQ

Anti-cheat Scoring — questions, answered.

Does anti-cheat work for all roles?

It's most useful for roles where coaching has high ROI — sales (script reads), engineering (LeetCode answers), customer success (canned scripts). For roles where coaching matters less (manual labor, hands-on trades), the score is informational.

What's the false positive rate?

Calibrated at <3% false positives at the default threshold. You can adjust the threshold per role — stricter for high-stakes roles, looser for high-volume top-of-funnel.

Will anti-cheat hurt candidate experience?

No — candidates don't see the scoring or know they're being monitored beyond the standard 'AI interview' disclosure. The signals are passive observations of the transcript and call audio, not active challenges. Real candidates feel like they had a normal interview.

What about ADHD, autism, or other cognitive variations that might look 'flagged'?

We deliberately don't score variation that's neurodivergent-typical (long pauses to think, atypical sentence structures, eye contact patterns). The thresholds are tuned to catch behaviors specific to coaching — not human variation.

Get started

Hire your next role with Raffi.

Start free with $25 in credits — no card required.

Get started free