Tavus (@tavus) on X

Tavus just released Griffin, an AI video model that passed what they're calling the video Turing test. The benchmark is straightforward: 48% of people who interacted with it live thought they were talking to an actual human. That's a massive jump from where the industry was before—previous systems topped out below 3%. The model also ranks first on NVIDIA's benchmark for full-duplex AI video, meaning it can handle real-time two-way video conversations without the lag that usually gives away AI systems. The company is positioning Griffin as the first "Human Interaction Model" or HIM. The naming suggests they're drawing a distinction from other AI video systems by emphasizing that this one is specifically built for natural human conversation rather than just generating video or handling one-directional tasks. Getting nearly half of test participants to genuinely believe they're talking to a person is the kind of capability threshold that matters—it's the difference between a tool that feels uncanny and one that actually works for real interactions. What makes this significant is the practical gap it closes. Video-based AI that can fool people in real time opens doors for customer service, sales calls, remote assistance, and other scenarios where human-like presence matters. The previous systems were so obviously artificial that they couldn't be deployed in contexts where the illusion needed to hold. Whether Griffin actually delivers on the hype in production environments is another question, but the benchmark results suggest the technology has crossed a real inflection point.