Tavus AI Model Griffin Tricks Half of Participants into Thinking They're Talking to Humans

Tavus, an AI startup, has released a new model called Griffin, which it claims is the first 'human-interaction model' capable of understanding and generating face-to-face conversations. The model not only understands language but also pays attention to facial expressions and pauses.

In testing, 54 participants engaged in one-minute video calls with the Griffin-Lite system, and 26 (48%) believed they were talking to a real person. In comparison, Tavus's previous system only convinced one of 41 participants that it was human.

Participants were told they would be having a video call with another person to discuss their expectations for the year, but were not informed until after the conversation ended whether the other person was actually a human or the model. Those who suspected the other person was not real typically became aware of this within 20 seconds.

Tavus claims that Griffin-Lite ranked first in the VideoFDB benchmark test by Nvidia, with a score of 3.83 in the generation dimension and 3.73 in the perception dimension. The model is also capable of full-duplex communication, allowing it to listen, see, and speak simultaneously.

The company completed a $40 million B-round funding in November 2025 led by CRV.