A startup says its AI passed as human in 48% of one-minute video calls, by its own test

Tavus, a San Francisco AI startup, says 26 of the 54 people who spent a minute on a video call with its new Griffin-Lite model believed afterward they’d been talking to a real person (Decrypt). That’s 48%. Its previous system faced 41 people and fooled exactly one, a pass rate of 2.4% (Crypto Briefing).

How Tavus ran the test matters as much as the number. Participants were told they’d be matched with another person for a one-minute call, and only at the end were they asked whether it had crossed their mind that their partner might not be real (Decrypt). Nobody on the call was told a bot might be on the other end. A community note on X flagged that the results aren’t independently verified and don’t follow a standard protocol, and Crypto Briefing notes that 54 participants is a modest sample.

Decrypt reports that Griffin listens, watches and talks at the same time, more like a phone call than a walkie-talkie, and its audio-to-video delay averages 0.43 seconds on NVIDIA H100 chips. On NVIDIA’s VideoFDB benchmark, Tavus says Griffin-Lite scored 3.83 on generation, against 3.92 for a human reference and 2.80 for the next-best system. For now it goes only to select trusted testers while Tavus works on safety measures and disclosure features (Decrypt).

Scammers already work on video calls. In January, North Korea-linked hackers used deepfakes on Zoom or Teams calls to pose as trusted contacts, then talked victims into installing malware disguised as an audio fix (Decrypt).

So if a video call ever asks you for money, a password or an install, I’d ask something no script would see coming. In 2025, Kraken’s security team flagged a suspected North Korean job applicant by asking spontaneous questions, like the names of local restaurants, and the candidate struggled (Decrypt).

A familiar face on a call proves less every month, so keep a question about lunch handy.