Tavus says its Griffin model fooled nearly half of people in a short live-video test

Tavus has introduced Griffin, an AI that conducts a video conversation continuously rather than waiting for a person to finish speaking. It watches facial expressions and gestures, hears tone and pauses, and generates its own voice, face, and reactions at the same time—skills that make interruptions, acknowledgments, and timing feel more natural. In Tavus’s 54-person, one-minute study, 26 participants thought they had spoken with a human, versus roughly 2% for the company’s previous system; Tavus calls that the first pass of a “video Turing test,” though the small company-run test is not a universal measure of human likeness. Griffin-Lite is only a limited research preview, but the result points to a shift: video AI may become far more useful for live assistance—and harder to distinguish from a real caller.