Know who is on the other side of the call.
Verified checks the person's camera, face and voice before the call, without joining it. It also looks for AI helper apps hiding from screen sharing, on Windows. Here is where it helps, and what we can prove so far.
Paste the numbers and limits into ChatGPT or Claude and ask your own questions.
For calls where it matters who shows up
- Hiring calls
- Client and sales calls
- Remote onboarding
- High-trust one-to-one calls
What is proven, and what is early
Proven today (in the product)
126 of 145
real takes Verified. One person, the founder, in 8 settings, on a Mac and a Windows laptop.
0 of 54
tricks got through, in the same scoring.
962 of 970
honest sessions cleared, and 0 of 6,308 tries with someone else's face got through. Recorded video of 70 people, tested offline. Partly in the product: the shipped display rule was not scored again on this test.
Early research (not in the product)
1,498 of 1,536
swap clips caught by a research model. Public research video, tested offline, not live.
66 of 66
live face swaps caught by a research model. One person, one Windows laptop.
What your team sees
- Verified
- The check passed.
- Can't confirm
- The check could not confirm it. The person is asked to try again, and you can ask about it on the call. It is not an accusation.
- Needs review
- Left for your team to look at.
What makes Verified different
-
We never join your call.
No bot in the participant list. Verified checks from outside the call, before it starts, and keeps an eye on it throughout. We chose this so the call feels like a normal call.
-
No special hardware, no visits, no eye scans.
Some identity checks need an eye scan or an in-person visit. Verified works with the laptop camera the person already has, on a Mac or a Windows computer.
-
Fair to the person being checked.
- Checked fresh on every call. Nothing is carried over.
- The wording never accuses anyone. It says "Can't confirm".
- The person sees exactly what happens first. A person can decline the check. They can still join the call with the host's link, and the host sees that they declined.
-
Simple for whoever runs the call.
The host sees a small badge next to each person in Meet, and a Verified button that opens the details.
Use case 1 of 5 Hiring
Someone else shows up for the interview
The person you screened is not the person on the call. You may find out after several rounds, or after the hire. That is hours of interviews, and a seat to fill again.
In testingVerified does not know who you expected to see. It compares the face on the call with the person's own camera on that computer.
Use case 2 of 5
A face swapped in by software
Live face-swap software puts a different face on the person in the call. On a normal webcam it can look convincing, so it can go unnoticed for the whole call.
Use case 3 of 5
An AI avatar or a replayed video
Instead of a live person, the call shows an AI avatar or a recording. On a small video tile it can look natural, and the time on the call goes to someone who is not there.
Use case 4 of 5
A face and a voice that don't belong together
The face on screen and the voice you hear come from different people. Small mismatches are easy to miss when you are listening to the answers.
In testingCatching someone else speaking off camera is not in the product yet.
Use case 5 of 5
A software camera playing prepared video
A software camera can feed any video into a call: a recording, or a face swap of someone else. To the call it looks like an ordinary webcam.
What it costs when the wrong person is on the call
When a trick gets through, a team spends hours of calls on the wrong person, sets up more meetings, and at worst makes a decision it has to undo, such as a hire.
Industry figures, not ours
In June 2022 the FBI warned of a rise in complaints about people using AI-altered video and stolen personal details to apply for remote jobs. In some of those interviews, it said, the person's lip movement did not fully line up with the voice.
FBI Internet Crime Complaint Center, public service announcement, June 28, 2022In one case announced in June 2025, the US Justice Department said the identities of more than 80 U.S. persons were used to get remote jobs at more than 100 U.S. companies.
US Department of Justice, press release, June 30, 2025
Where Verified fits
Today's tools usually check a person once, for example at sign-up. Verified checks the person on the live call, before it starts and throughout, without joining the call.
At sign-up
ID and selfie checks
During a test
Interview proctoring
On the live call
Verified, before it starts and throughout
How we test
We test on recordings our team never saw while building.
A separate person who didn't build the check scores it.
We write down what counts as a pass before we test. Failed tries stay on this page.
What we can't show yet
- Most tests so far are one person, the founder, on one Mac or one Windows laptop. The numbers are small, and we say how small beside each one.
- Windows runs fewer checks today. The face check is not ready on Windows yet.
- No single check catches everything. Some tricks are left to a different check, and some are not covered yet.
- A harder voice test, with face-swap tools the system never trained on, did not pass. It scored 98 of 120 (81.7%; likely range 73.6 to 88.1).
- The face check pauses while the call tile is hidden, for example during a screen share, and it needs the desktop app.
- Real people can be wrongly flagged, mostly when the video is very low quality (8 of 970 in our tests).
- Not everyone passes on the first try; the app explains what to change and offers a redo. Real-world first-try rates will come from the beta.
- Verified does not replace an ID check. Intel Macs are not supported.
- Not in the product yet: the head-and-hands swap models and a check for a second voice near the person on the call.
- Results are advice for your team, not a verdict on a person.
Availability
Closed beta for 20 companies. Works with Google Meet, on Mac computers with Apple silicon and Windows PCs. Contact us to join.
Most of our numbers were measured between September 30 and October 8, 2026 and scored by a person who did not build the check. One set (126 of 145) is a second scoring of earlier recordings, and the research models were checked on October 9, 2026. Raw counts sit beside every percentage. Where we say "we are confident it is at least", it is the lowest rate the number of tries can support with 95% confidence.
Last updated
Keep reading
How Verified works before a call