The claim
On October 1, the San Francisco startup Tavus announced Griffin, a real-time video-to-video model it calls the first "Human Interaction Model." Earlier conversational AI works like a walkie-talkie: it waits for you to finish, thinks, then replies. Griffin is full-duplex. It watches, listens, and generates video and speech at the same time, so what it sees can change what it says while it is saying it. It handles interruptions, laughs when the moment calls for it, and keeps the video frame moving throughout.
Then came the headline. Tavus says Griffin is the first model to pass a real-time video Turing test. In a blind study, 48% of participants who spent one minute on a video call with a Griffin-powered system believed their partner was a real person.
The fine print
The study is Tavus's own. Fifty-four participants, one minute each, recruited through an unnamed platform, with no outside evaluator in the loop. Tavus's previous stack fooled 1 of 41 people, a comparison that sets up a dramatic before-and-after, except both numbers come from the company that benefits from them. There is no agreed definition of a "video Turing test," so "passing" means clearing a bar Tavus drew.
The independent number is better, though thinner. On NVIDIA's VideoFDB benchmark, Griffin scored 3.83 on generation against a 3.92 human reference, and 3.73 on perception against 4.20 for people. Genuinely close to human on generation. Further off on perception. Tavus also claims a 37% edge in real-time reactivity over the next-best system, a metric described only in its own research report.
What actually changed
Set the Turing-test marketing aside and the architecture is the interesting part. Tavus's previous systems chained separate pieces: one model rendered the face, another read the room. Griffin puts perception, decision, and generation inside one model, so a gesture it sees mid-sentence can reshape the sentence. That is a real change in how conversational video gets built, whether or not the Turing-test framing survives.
The part Tavus is not selling
Note what Tavus did not do: ship it. Griffin-Lite is a research preview for select testers. The company says a fuller version arrives after it addresses safety concerns, and it is holding back customer access until it builds disclosure features, ways for the person on the other end to know they are talking to a model. That is the most honest sentence in the announcement. A system that passes as human on video calls is, by definition, a system that can deceive people on video calls.
The listed use cases are sensible: tutoring, practicing difficult conversations, camera-based tech support. The unstated ones are why the gate exists.
One minute is a short test
Stay on the line for ten minutes and most systems show their cracks; Tavus has not shown us that call. Griffin deserves attention because full-duplex video conversation is a genuine new capability. It deserves skepticism because the headline number is self-reported. Judge it when strangers can try it, not when the vendor runs the test.
Sources
- [1] businesswire.comRead source
- [2] the-decoder.comRead source
- [3] cellcog.aiRead source
- [4] cryptobriefing.comRead source