Tavus's Griffin: an AI Avatar People Mistook for a Human

The startup Tavus has unveiled Griffin, the first model in a new category called Human Interaction Model, an AI tuned for live conversation rather than text. It reads the other person's facial expressions, tone of voice, and pauses, and responds with real-time video. In a Tavus study, 48% of participants mistook Griffin for a live human after a one-minute video call. Earlier systems topped out at just 2% in similar tests. In an independent Nvidia test measuring how human a video call feels, Griffin scored 3.83 points. Live humans scored 3.92, and the best prior model scored 2.80.

For agents that need to sound and look alive, in support, sales, training, this sets a new bar for quality. The gap with a live human in the Nvidia test is small, 3.83 versus 3.92, and this is not a marketing claim, it is someone else's measurement. But you cannot try it yet. Only a lightweight Griffin-Lite is available, and only to select testers, while the full version is being held back over safety concerns. So it is too early to build this into a real product, and I would watch for updates rather than factor it into any plans right now.