Tavus releases Griffin and discloses small-sample real-time video test results

Article

Tavus releases Griffin: 54 people video chatted for 1 minutes, 26 people thought the other party was a real person

Beating AI News Brief,AI the video company Tavus released Griffin, focused on real-time face-to-face communication. Tavus calls it the first Human Interaction Model(Human Interaction Model). Griffin can see and listen while speaking, responding in real time to the other party’s facial expressions, movements, and pauses. If the user interrupts, it can stop immediately.

Tavus recruited 54 participants from the United States and Europe to video-call Griffin-Lite for 1 minutes, without telling them in advance that the other party was AI. The result: 26 people, or 48%, believed they had just been chatting with a real person. By comparison, Tavus’s previous-generation Phoenix-4.5 achieved only 1/41 in a similar test, meaning 2.4%of participants made the wrong judgment.

Tavus therefore calls Griffin the first model to pass the “real-time video Turing test.” However, the test was organized by Tavus itself, involved only 54 participants, and has not yet been independently replicated at a larger scale.

Griffin-Lite also ranked first in NVIDIA’s VideoFDB full-duplex audiovisual evaluation. Tavus reported a score of 3.83/5, close to the human reference score of 3.92. The highest previous public baseline,Gemini 2.5+ Anam, scored 2.80.

Griffin-Lite is currently available only to a small number of testers and has not yet been formally offered to ordinary customers.

Original article link https://m.theblockbeats.info/flash/369981

Key points

Tavus released Griffin and called it the first Human Interaction Model(Human Interaction Model).
Griffin can respond in real time to the other party’s facial expressions, movements, and pauses, and can stop immediately when the user interrupts.
54 participants from the United States and Europe held a 1-minute video call with Griffin-Lite, during which 26 people thought the other party was a real person, accounting for 48%。
previous-generation Phoenix-4.5 achieved 1/41 of participants made the wrong judgment, accounting for 2.4%。
Tavus said Griffin is the first model to pass the “real-time video Turing test.”
Griffin-Lite scored NVIDIA in the 3.83/5 VideoFDB full-duplex audiovisual evaluation, and public information says it ranked first.
The human reference score was 3.92 The previous highest public baseline, Gemini 2.5+ Anam, scored 2.80。
Griffin-Lite is currently available only to a small number of testers and has not yet been formally offered to ordinary customers.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Simply put, Griffin is a AI designed for real-time video conversations. The reported tests show that some participants could not tell that they were facing a AI; however, the test was organized by Tavus itself and was small in scale.

Why it matters to readers

It indicates that real-time video AI may be closer to human communication in handling speech, facial expressions, movements, and interruptions.

This is not an ordinary prerecorded digital human, but one that responds instantly during video conversations based on the user's reactions AI system.

Risks and unknowns

The test was organized by Tavus itself, with a sample of only 54 people.
There is currently no larger-scale independent replication.
The report did not provide complete testing methods or details on participant screening.
Griffin-Lite has not yet been officially made available to ordinary customers.
The claims of being the “first to pass the real-time video Turing test” and the evaluation rankings are based primarily on Tavus's statements.

Tavus releases Griffin and discloses small-sample real-time video test results

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center