Tavus says its Griffin model passed the video Turing test

In 1950, Alan Turing asked a simple question: can a machine convince the person it’s talking to that it’s human? On October 1, AI company Tavus took that question to video calls.

It ran one-minute video calls between 54 people and its new model, Griffin, without telling them they were talking to AI. 26 of them, or 48%, thought they’d spoken to a real person. With the company’s previous system, that number was just 2.4%, one person out of 41.

The Turing test’s original name is the “imitation game.” In a 1950 paper, Alan Turing turned the question “Can machines think?” into a game: an interrogator chats by text with two parties, one human and one machine. If they can’t tell which one is the machine, the machine passes.

The text version was passed in 2025: in a UC San Diego study, interrogators picked GPT-4.5 as the human in 73% of five-minute chats. The video version is much harder, because the face, voice, expressions and timing all have to be convincing too.

What sets Griffin apart is that it listens, watches and talks at the same time. Most systems today take turns: the user speaks, the system listens, then it answers. Griffin talks the way people do. It can change its answer mid-sentence, and it waits when the user pauses to think.

In one demo, it worked an object shown on camera into the story it was telling. And it generates the whole face, expressions and movement live from a single photo.

The “passed the video Turing test” claim is Tavus’s own, and it’s worth reading carefully. The company ran the study itself, with no outside evaluator. Just 54 people and one-minute calls.

Participants were told they’d be talking to another participant, so they were already expecting a human. With a group this small, the real rate could be anywhere between 35% and 61%. That’s why a community note was added to the announcement on X.

NVIDIA’s video conversation benchmark (out of 5)

A real human: 3.92 Griffin: 3.83 Next best AI: 2.80

On NVIDIA’s benchmark for real-time video conversation, Griffin came close to a real human and left other systems far behind. The delay between speech and video averages 0.43 seconds.

Tavus hasn’t released the model publicly. Only select testers have access while the company builds safeguards and “you’re talking to AI” disclosures.

The reason is obvious: an AI that passes for human on video could be a patient tutor or support agent, but also a very convincing scammer.