Tavus's Griffin passed as human in 48% of video calls

San Francisco startup Tavus unveiled Griffin on October 1, the first of what it calls a Human Interaction Model. It listens while people talk, reads expressions, gestures and pauses, generates a real-time talking human video, and can interrupt like a person.

Participants recruited through an independent platform were told they would have a one-minute video chat with another participant about what they were looking forward to this year. Their partner was a Griffin-Lite persona. Asked afterwards, 26 of 54 (48%) believed they had talked to a human, against 2.4% for Tavus's previous system, Phoenix-4.5.

Griffin merges real-time perception, conversation and expressive video into one full-duplex video-to-video system, generating 720p video in 320 ms chunks with 0.43 seconds of audio-to-video latency on Nvidia H100s, below what people notice in conversation.

It enables personalized video at scale for support, sales and education, but also raises the risk of video fraud and deepfakes. Tavus limits access to select testers pending safety evaluation.

The calls lasted one minute on one topic with 54 people, and the results are Tavus's own; independent replication is pending.

Alan Turing proposed in 1950 that a machine able to convince a person it is human in conversation could be said to think. Text chatbots have begun passing short versions; video is far harder because people notice tiny inconsistencies in faces and timing.

Face-to-face AI could handle customer support, sales calls, tutoring and pre-appointment health screening at scale, with each conversation personalized. The same capability makes video impersonation cheaper, which is why Tavus is gating access while it evaluates safety.

Disclose when customers speak with an AI assistant, verify critical requests such as transfers through a second channel even if they arrive by video, and warn staff about video calls impersonating executives.