Griffin, an AI model from Tavus, generated realistic video calls that convinced 48% of test participants they were talking to a real person. The system takes a reference image and generates the entire scene in real time—body movements, gestures, shadows, background, and all—while cloning the subject's voice from just 10 seconds of audio.
The model understands and produces face-to-face conversations in real time, capturing not just words but also facial expressions and natural pauses. In a test with 54 participants conducting one-minute calls, nearly half believed they were speaking with a human. That's a dramatic leap from Tavus's previous system, which fooled only 2.4% of users.
For now, the technology remains restricted to selected testers. The progress shows how quickly synthetic video is becoming harder to distinguish from reality—a shift with significant implications for authenticity online.