Tavus unveils Griffin AI model capable of passing video Turing test

Tavus, a California-based AI startup, has introduced Griffin, a new Human Interaction Model (HIM) designed for real-time, full-duplex video communication. The model is capable of listening, observing, and reacting simultaneously, allowing it to nod, handle interruptions, and adjust facial expressions in intervals of less than a second.

In a study conducted by the company, 26 out of 54 participants (48%) believed they were speaking to a real human during a one-minute video call. This represents a significant increase from the company’s previous system, which had a success rate of only 2.4% (1 out of 41 participants). Tavus describes this achievement as passing a real-time video Turing test.

While the results are notable, observers point out that the test was conducted by the company itself and involved short, one-minute interactions where participants were not initially told they might be speaking to an AI. Due to the potential for misuse in deepfake scams, the company has not yet released the model to the general public, opting instead to conduct further testing and develop disclosure mechanisms to identify the artificial nature of the avatars.