Griffin from Tavus Passes Video Version of the Turing Test

Startup Tavus introduced the Griffin model, which the company calls the world's first system to pass a video version of the Turing test. In the experiment, 48% of participants believed they were communicating with a real person.

Model Capabilities

Griffin can listen, watch, and speak simultaneously. The system nods and laughs while the other person is still talking, incorporating what it sees through the camera into the conversation.

• Generates a full frame with hands and gestures, not just the face
• Responds with a delay of only 0.43 seconds
• Outperformed all published baseline systems on the NVIDIA VideoFDB benchmark

The testing results are based on a sample of 54 people, of whom only 26 tested the lightweight version, Griffin-Lite. By comparison, only 2.4% did so for Tavus's previous system.

Limitations and Security

Tavus acknowledges that Griffin can convince a person that they are not interacting with an AI. Therefore, the model is currently available only to select testers. As for protective measures, they promise disclosure features that are still under development.