48% thought Tavus’s AI was human

Tavus has put a new name on a familiar unease: Griffin, its first Human Interaction Model, or HIM. The company says the system is built to understand and generate face-to-face, real-time interaction, not just text or canned voice replies. In a live study, 48% of participants who spoke with Griffin believed they had been talking to a real human after a one-minute video call.

That number is the headline, but the more interesting part is how Griffin works. Tavus says the model listens, watches, and reacts to expressions and pauses while it talks, using a unified video-to-video system that combines perception, conversational modeling, and expressive human-like video. The company contrasts that with older systems that wait for the user to stop, think, and then respond. Griffin, by design, keeps reassessing the exchange in sub-second mini-turns.

The result is a system that can backchannel, interrupt, pause, or react to visual context without waiting for a clean turn to begin. Tavus says Griffin can nod while someone is speaking, stop when cut off, and respond to things it sees, not just what it hears. It also claims the model can handle full-scene generation in real time from one reference image, including the chair, shadows, and background, not just the face.

Tavus is calling the preview version Griffin-Lite and says it is available today to a select group of early testers. A wider release of a more powerful model is supposed to follow. The company also says Griffin took first place in NVIDIA’s independent test of face-to-face AI and was 37% ahead of the next best AI at reacting in the moment.

The pitch here is obvious: if a machine can keep up with your pauses, expressions, and interruptions, it starts to feel less like software and more like company. That is either the future of computing or the world’s most elaborate way to make people talk to a face on a screen.