How interview integrity tools are tested, and what good evidence looks like

This page is for buyers comparing vendors. It shows what interview integrity vendors publish today, what a credible test report contains, why fairness across groups belongs in it, and how we publish our own. For the list of questions to put to a vendor, see why you should ask for the false-alarm rate.

What vendors publish today

Our list of interview cheating detection tools covers 15 products by approach.

For several other vendors we checked, including interview, identity and background-check providers, we did not find a public accuracy or false-alarm figure as of October 29, 2026. That does not mean none exists; it means a buyer has to ask. Most tools share a headline number. Few publish the test behind it.

Why a headline number is not enough

A single accuracy figure hides the two numbers a hiring team needs: how often the check catches a real problem, and how often it flags an honest person. It also depends on the mix. If almost every session is honest, a check that flags nobody scores very high "accuracy" and catches nothing.

Results also move between the lab and real life. Researchers who built the Deepfake-Eval-2024 benchmark found open-source detectors lost ground on deepfakes collected from the real world, with "AUC decreasing by 50% for video, 48% for audio, and 45% for image models" compared with older benchmarks. The study is a preprint, but the lesson holds: ask how the test data was gathered, and when.

What good evaluation looks like

1. Both numbers, as counts. Problems caught and problems missed; honest people cleared and honest people flagged.
2. Real conditions. Live calls or recordings, which devices, and what else was open on the computer.
3. Data the builders never saw. If a team tunes a check on the same data it scores, the score overstates real use.
4. A confidence bound. A small sample needs a range, not just a percentage.
5. Results across groups. The same check can treat demographic groups very differently.
6. Dates and re-runs. When the test ran, and what happened when an earlier run went wrong.
7. Independent testing where it exists. NIST runs ongoing face recognition evaluations that report error rates at set thresholds, for example: "FNMR is the proportion of mated comparisons below a threshold". The FIDO Alliance certifies face verification and says it will "test bias levels, and measure biometrics across different demographics".

Fairness: the adverse-impact lens

In hiring, a false alarm is not spread evenly. NIST's 2019 study of face recognition found large differences in false positives between demographic groups.

The differentials often ranged from a factor of 10 to 100 times

US hiring rules already have a yardstick. Under the Uniform Guidelines, a selection rate "less than four-fifths (4/5) (or eighty percent) of the rate for the group" with the highest rate is generally taken as evidence of adverse impact, though the same section notes that some "differences are based on small numbers and are not statistically significant". A check that flags one group more often can push a hiring process toward that line. So a test report should show results across groups, not just one overall figure.

Small samples need a range

Zero false alarms in 30 tries is a good start, not proof of zero. A 1983 paper in JAMA, "If nothing goes wrong, is everything all right?", set out the rule of three: with no events in n tries, the true rate is still likely below about 3 in n, at 95% confidence.

This is why we show the lowest rate each result supports next to the count. How to read a small sample in more detail is in the false-alarm guide.

How Verified Call publishes its tests

Most tools share a headline number. We publish how we test, false alarms included. For each check, our interview cheating detection test results show what we did, the count, the false alarms, the sample, the conditions, the test date and a 95% confidence bound. Results from recordings say so, and so does research that is not in the beta.

• Hidden AI tools. Found in "30 of 30 tries, for each of 4 tools", at least 90.5% for each tool (95% confidence); 0 of 30 false alarms with no AI tool open, and 0 of 30 with everyday apps open. Our own testing, October 4, 2026.
• Camera software and direct cameras. "29 of 29 tries each, for 9 kinds of camera software and 7 direct cameras", at least 90.2% for each one. Our own testing, September 30, 2026.
• Live calls. 20 of 20 live calls read right (10 direct cameras, 10 face swaps sent through camera software), October 2, 2026.
• Someone else's face. 0 of 3,990 sessions got through (2,855 flagged, 1,135 ended "Can't confirm"); the real person was cleared in 962 of 970 sessions, 99.2%, at least 98.5% (95% confidence). Public research video of 70 people, tested offline on recordings our team never saw while building, October 1, 2026.

Failures stay on the page too. The camera result notes: "An earlier run read some direct cameras wrong. We fixed it and tested again. Both results are on record." Why we do it this way is in the founder's post, why we publish how we test, false alarms included.

Verified Call checks the person on the call, before it starts and while it runs, without joining. It checks for hidden AI answer tools, camera software, face swaps, deepfakes and AI avatars sent through camera software, and someone else answering. It checks that the voice on the call matches the person who did the check.

It works inside the video calls you already run. No bot in the participant list. It checks on the person's own computer, so it sees what a video call cannot carry: a tool hidden from screen sharing, or a face sent in through camera software. Each call is checked fresh. The host sees a plain result, with proof beside any flag, and a person can decline the check and still join with the host's link; the host sees that they declined.

What to ask any vendor

Three questions cover most of it. How often does it flag an honest person, as a count? What was tested, on whom, live or recorded, and when? Can we see results across groups? The full list of eight is in the false-alarm guide, and industry numbers are on the statistics page.