San Francisco startup Tavus unveiled Griffin, which it calls a Human Interaction Model – an AI that can see, hear, speak and move like a person during a live video call. In a one‑minute test, 48 % of people said Griffin felt like a real human. Earlier AI video avatars rarely passed the human check, with rates around 2 %.
How Griffin was measured
The company ran an independent test with Nvidia that scored how "human" the AI seemed during a direct audio‑video conversation. Griffin earned a 3.83 rating on a 5‑point scale, while real humans scored 3.92. The previous best AI model reached only 2.80.
Where the tech could appear
Tavus says Griffin could be used for tutoring, rehearsing tough talks, or as a camera‑based tech‑support assistant. A lighter version called Griffin‑Lite is being offered to a small group of research testers, with a fuller version planned after safety concerns are addressed.
Why it matters
If Griffin can convincingly act like a person, everyday users could practice conversations or get help from a realistic virtual assistant without needing a human on the line. However, the technology is still in a research preview, so wide public access and safety safeguards are not yet in place.
