Tavus says its Griffin AI passed a real-time video Turing test

Tavus, a San Francisco startup, says its new video model Griffin has passed a real-time video Turing test. In the company's own study, 26 of 54 people, or 48 percent, believed they had just talked to a real human during a one-minute video call. The announcement came on 1 October 2026.

What happened

Tavus calls Griffin a Human Interaction Model: one system that watches and listens while it talks, then produces voice, face, gaze and gestures at the same time. Most live AI avatars work like a relay race, where speech is transcribed, a language model writes a reply, and other tools turn it into a voice and a face. Every handoff adds delay and loses cues such as tone or what the camera sees. Griffin decides many times per second whether to speak, nod, interrupt or stay quiet.

For the test, participants were told they would chat with another study participant about what they look forward to this year. Their partner was actually Griffin-Lite, a smaller preview version. Only at the end were they asked whether they had suspected an AI. Tavus reports that its previous stack convinced just 1 of 41 people under the same protocol. Tavus also cites NVIDIA's VideoFDB benchmark, where Griffin-Lite scored highest on both tracks.

Griffin-Lite is not a product. Only a select group of trusted testers can use it as a research preview, and Tavus says a wider release depends on disclosure and safety work.

Why it matters

Realistic AI video is moving from pre-rendered clips to live conversation. That opens useful doors, like tutors, practice partners and support agents that react to your face. It also makes impersonation and scams much easier, since a video call has long felt like a safe proof that someone is real.

The number deserves care. It comes from the company itself, the sample is small, the call lasted one minute, and people were not told that an AI might be on the other end. That is not the classic Turing test. Still, the jump from about 2 percent to nearly half under one protocol is hard to ignore.

Dany's take

I like that Tavus is holding the model back, but a pause only helps if disclosure becomes normal. If a face on a call can fool half of a small group, we need clear labels, and we should not rely on gut feeling. I will watch for independent replications with longer calls and bigger groups. Until then, treat 48 percent as a strong signal, not a settled fact.

Source: Tavus: Griffin, the First Human Interaction Model.