Tavus says its Griffin video model passed a video Turing test, with 26 of 54 testers taking it for a real person

Tavus, a San Francisco AI video company, announced Griffin on 1 October. It calls Griffin its first “Human Interaction Model”: one video-to-video system that sees, listens and generates a talking face in real time. Tavus claims “Griffin is the first model to pass the real-time, video Turing test.”

The claim rests on Tavus's own study. Participants were told they would join a one-minute video call with another participant, but their partner was a research preview called Griffin-Lite. “Of the 54 participants who spoke with Griffin-Lite, 26 believed their partner was a real person,” the company says, which is 48%. For its previous model, Phoenix-4.5, the figure was 2.4% of 41 participants.

Tavus also cites Nvidia's VideoFDB benchmark, which it says Nvidia ran independently. By Tavus's account, Griffin scored 3.83 on generation, against 3.92 for humans and 2.80 for the next-best system. No outside group has repeated the study.

Griffin is not on Tavus's platform, and Griffin-Lite is limited to selected trusted testers. The company says the qualities that make such models work also “allow them to deceive a human into believing it is not AI”, and that it is building safe-disclosure features before a wider release.

• Claimed Tavus says Griffin is the first model to pass a real-time video Turing test. Tavus
• Claimed In Tavus's own one-minute study, 26 of 54 participants (48%) believed Griffin-Lite was a real person, against 2.4% of 41 for Phoenix-4.5. Tavus
• Claimed Tavus says Griffin scored 3.83 on generation in Nvidia's VideoFDB benchmark, against a 3.92 human reference. Tavus, via Business Wire
• Confirmed Griffin-Lite is limited to selected testers while Tavus builds disclosure and safety features. Tavus