Photo: Steve A Johnson / Pexels
The San Francisco AI lab says nearly half of blind study participants mistook its new model for a real person on a video call
Tavus wants you to second-guess your next video call. The San Francisco AI research lab unveiled Griffin on October 1, 2026, and says it is the first model to pass a video Turing test.
In a live blind study, 48% of participants who chatted with a Griffin-powered system believed a human was on the other end.
What Tavus actually built #
Tavus calls Griffin the world’s first Human Interaction Model, or HIM. Think of it less as a talking avatar and more as a full conversation partner that watches, listens, and responds in real time.
Technically, Griffin is a full-duplex video-to-video system. Full-duplex means both sides can talk and react at once, the way people interrupt, nod, and overlap in a normal chat. Earlier systems tended to work more like walkie-talkies, where one side waits for the other to finish.
The model pulls together perception, generation, movement, and conversational modeling into a single architecture. That matters because Tavus previously chained separate systems together: Phoenix handled rendering, while Raven handled perception.
The numbers behind the claim #
The headline result comes from a blind study with 54 participants. Each person spent one minute on a video call with a Personified Application Layer, or PAL, running on Griffin.
Nearly half of them, 48%, walked away thinking they had spoken with a genuine human. Tavus’s previous system managed a pass rate of just 2.4%, fooling 1 out of 41 participants.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Tavus also pointed to NVIDIA’s VideoFDB benchmark. Griffin scored 3.83 on generation, against a human reference score of 3.92.
The company says Griffin beat the next-best published system by 37% in real-time reaction metrics.
The announcement came from CEO Hassaan Raza and Head of Research Ioannis Patras. For now, access is limited.
Tavus is offering a research preview called Griffin-Lite to selected testers. The company plans to release a fuller version after it addresses safety measures.
Why the study design deserves a closer look #
A 48% rate sits close to a coin flip, which is arguably the most meaningful threshold: if people cannot do better than chance at spotting the machine, the machine is effectively indistinguishable.
The flip side is that 54 participants is a modest sample, and one minute is a short window.
Where Griffin fits in Tavus’s business #
Griffin builds on Tavus’s existing PAL platform. According to the company, more than 150,000 developers and businesses have used PAL.
For those customers, the appeal is straightforward. Any business that relies on video interactions, from customer support to onboarding, could put a far more convincing digital representative in front of people.
What this means #
Tavus is holding back the fuller version until safeguards are in place. A system that half of people cannot tell from a human is useful for customer service, and equally useful for anyone hoping to impersonate someone convincingly.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our