Tavus Avatar Fools Nearly Half of Callers in Test

Tavus says its new Griffin avatar passed as human to 48% of people on a one-minute video call, a dramatic jump from prior systems that barely registered a few percent.

Tavus has unveiled Griffin, a system the company describes as its first Human Interaction Model, designed to hold live video conversations that look and feel like talking to an actual person. Rather than just generating a static face or lip-synced clip, Griffin is built to read and respond to a caller's facial expressions, tone of voice, and physical gestures in real time, adjusting its own behavior as the conversation unfolds. The headline figure comes from testing Tavus ran on its own model: after a one-minute video call, 48 percent of participants reportedly believed they had been speaking with a real human rather than an AI avatar. For comparison, the company says earlier-generation systems only managed to pass as human in roughly 2 percent of similar tests, suggesting a sharp jump in how convincing these conversational avatars have become in a short span of time. It's worth noting that this result comes from an internal Tavus study rather than an independent audit, so the figure reflects the company's own framing of Griffin's performance rather than third-party verification. Still, even as a self-reported benchmark, a near-coin-flip success rate at fooling people face-to-face marks a notable milestone for real-time AI avatars, an area that has mostly been judged on scripted demos rather than live back-and-forth interaction. The jump also raises familiar questions that tend to follow any leap in synthetic media realism: how such avatars might be used in customer service, sales, or virtual assistants, and how easily the same underlying tech could be repurposed for impersonation or fraud if convincing real-time video becomes cheap and widely accessible. Tavus pitches Griffin as a product for legitimate interactive use cases, but a model that can pass as human in casual conversation inevitably invites scrutiny over consent, disclosure, and misuse. More detail on the study and the model itself is available in the original report.