Tavus Launches Griffin: Half of Video Call Participants Didn't Spot the AI

On October 1, US AI company Tavus released the real-time video model Griffin. It calls itself the first “Human Interaction Model” (HIM): in one-minute video calls, 48% of participants believed they were talking to a real person.

How the 48% Was Measured

The number comes from Tavus’s own blind test. 54 participants each had a one-minute video call with Griffin-Lite. Before the call, they were told the person on the other end was another participant.

Afterward they were asked: do you think you were just talking to a real human? 26 said yes — 48%. By contrast, Tavus’s previous digital-human stack (Phoenix-4.5 plus Sparrow-2 plus Raven-1) fooled 1 of 41 people — 2.4%.

Tavus CEO Hassaan Raza and research lead Ioannis Patras co-wrote the launch on the official blog. They didn’t disclose recruitment channels or bring in third-party evaluators. With only 54 participants, the conclusion deserves discounting. Even so, the leap from 2.4% to 48% signals real progress on “feeling human.”

How Griffin Differs From Past Digital Humans

Old digital humans were pipelines: speech recognition, language model, speech synthesis, facial animation — running one after another. The result: it finished a sentence before starting to process a reply, with awkward pauses in between.

Griffin folds those stages into one model. Video-to-video, full duplex: its eyes keep watching and ears keep listening while it talks. Interrupt it and it stops instantly; when you pause, it judges whether you’re hesitating or done, then decides whether to jump in or wait. It drops “mm-hm” acknowledgments into your pauses, with nods, head tilts, and gestures following the conversation’s rhythm.

Tavus’s example: Griffin can watch you solve a Rubik’s Cube while teaching, offering hints along the way. In other words, the model sees the picture — not just the transcript.

The Technical Specs

Per Tavus’s published numbers: 720p output at 25fps, average audio-video latency of 0.43 seconds, running on NVIDIA H100s. Generation needs one photo and ~10 seconds of audio to build a digital double.

On NVIDIA’s VideoFDB benchmark (for full-duplex audio-video dialogue), Griffin scored 3.83 in the generation track against a human baseline of 3.92, with the runner-up system at 2.80. Perception track: 3.73 versus 4.20 for humans. Its median response latency is 2,232 ms, though — slower than open-source MiniCPM-o.

Why You Can’t Use It Yet

Griffin-Lite currently goes only to a handful of trusted testers as a research preview. No public API, no pricing. Tavus’s line: no public release until safety mechanisms are ready.

The reason is direct. A 48% deception rate breaks the assumption that “a video call verifies you’re talking to a real person.” Last year a Hong Kong company lost $25 million to an AI deepfake video-conference scam with everyone impersonated. The tooling bar keeps dropping: one photo plus 10 seconds of audio is enough for a real-time interactive digital double.

Tavus says it’s building disclosure mechanisms: the system should proactively say the other party is AI. Until that’s ready, it stays in research preview.

The End of “Seeing Is Believing”

Griffin’s significance isn’t another more-realistic digital human. It forces a long-standing problem into the open: when AI can listen, watch, speak, and react in a video call — handling interruptions and pauses like a person — traditional trust signals (a glance, a few words) stop working.

Tavus released a short film starring Vanessa, a digital human running on Griffin. It calls this “human computing”: the machine disappears, and you talk to it like a friend. That sounds gentle — but the same capability in a scam ring’s hands is another story.

Two things to watch next. First, what Tavus’s safety disclosure mechanism looks like. Second, whether new standard practices emerge for identity verification on video calls. Until both answers land, Griffin stays in research preview.