Nearly half of test subjects mistook Tavus' AI video avatar for a real person on a one-minute call

Tavus introduced a research preview of Griffin-Lite, reporting that 26 of 54 participants (48%) believed they had spoken with a real person after a one-minute live video call. The result underscores t

Griffin, unveiled by Tavus on October 1, is a “Human Interaction Model” that receives the other person’s video and voice in real time and converses with them. In the company’s Griffin-Lite research preview experiment, 26 of 54 participants, or 48%, judged the other party to be a real person after a one-minute video call. The figure shows that even a brief conversation can already destabilize judgments about identity.

Mistaken-identity rate far higher than the comparison system

In a comparison conducted in the same way, only 1 of 41 people (2.4%) said Tavus’s previous Phoenix-4.5-based system was a real person. Compared with Griffin-Lite’s 48%, this suggests that interaction timing—such as turn-taking, interruptions, gaze, and facial expressions—has a greater influence on creating a human-like impression than video generation that merely matches lip movements.

What did the experiment measure?

Participants recruited in the United States and Europe were told they would be connected with another participant and talked for one minute about what they expected from the year. The other party was actually a Griffin-Lite-based PAL that generated a face, voice, and responses in real time. Only after the call were they asked whether they had considered that the other party might be AI and what they believed the other party’s identity to be, and all participants were informed that it had been AI. This measured perceived realism; it was not a test proving the accuracy of the responses or reliability equivalent to that of a human.

Remaining challenges in service design

For services using video agents, such as customer support, education, and counseling, increasing naturalness alone is not enough. They should clearly disclose whether the agent is AI at the start of the conversation and establish a path for handing incorrect answers or sensitive requests over to a human. Tavus introduced Griffin-Lite as a research preview available to a limited group of early testers and said it plans to release a more powerful version more broadly later.

Source: Official Tavus Griffin announcement · The Decoder report