False Positive Rates in AI Proctoring: What HR Leaders Should Demand
Every proctoring vendor claims to catch cheaters. Almost none will tell you how often they accuse innocent candidates.
That number is the false positive rate. It is the single most important metric in interview fraud detection, and most vendors will not publish it. Ask why.
A false positive is not a rounding error. It is a qualified candidate your company rejected for cheating they did not do. At volume, a 2% false positive rate on 10,000 interviews is 200 good candidates thrown away. Some of them will post about it. One of them will call a lawyer.
Demand a published number
If a vendor will not publish its false positive rate, assume it is bad. Publishing the number is cheap for a vendor with a good one and expensive for a vendor with a bad one. That is exactly why the demand works.
The number should be specific. "Under 0.4% across high-volume customers, published quarterly" is a claim you can hold someone to. "Industry-leading accuracy" is a sentence designed to end the conversation.
Ask how the rate is measured. Against what ground truth? Over what sample? Updated how often? A rate measured once on a friendly dataset is marketing. A rate tracked quarterly across live customers is operations.
RightMatch publishes its false positive rate quarterly. It is tracked at under 0.4% across high-volume customers. Hold us to it.
Demand evidence per flag, not just a score
A score is an accusation without an argument. "Fraud risk: 92%" tells the recruiter nothing they can act on and nothing they can defend.
Demand a timestamped, explainable evidence trail for every flag. What signal fired. When. At what confidence. What the candidate was doing at that moment. A flag that reads "AI-generated voice detected at 0:14, 1:32, 4:48, high confidence" is reviewable. A flag that reads "suspicious" is a lawsuit waiting for a plaintiff.
The evidence trail is what survives an HR legal review. Your counsel does not care about the model. They care about the record: what was observed, when, and who made the final call.
Demand a human in the loop
No system should auto-reject a candidate on fraud signals alone. The composite score should trigger quarantine and review, not a verdict. The recruiter decides. The system advises.
This is also where single-signal discipline matters. A VPN flag, a tab switch, an unusual pause: each is a signal, never a verdict. Auto-block should require multiple corroborating signals and a threshold you set per role. If a vendor's system blocks on one signal, walk away.
The five questions for every vendor
- What is your published false positive rate, and how often do you update it?
- What does the evidence trail look like for a single flag? Show me one.
- Can a single signal trigger a block, or does it require corroboration?
- Who makes the final decision: the system or the recruiter?
- What is logged, for how long, and can my legal team audit it?
If a vendor answers all five cleanly, you have a partner. If they dodge the first one, you have your answer.
The line: A vendor that will not publish its false positive rate is telling you the number. Demand the rate, the evidence trail, and a human in the loop, or walk away.