Tavus Unveils Griffin AI Avatar That Fooled Nearly Half of Video Callers
AI startup Tavus has introduced Griffin, a real-time conversational model that 48 percent of test subjects mistook for a real person during a one-minute video call.
Tavus Introduces Griffin Interaction Model
San Francisco-based startup Tavus has launched Griffin, an AI system it describes as the first "Human Interaction Model" (HIM). Built to conduct fluid, face-to-face video conversations in real time, Griffin is designed to bridge the gap between static conversational bots and realistic video presence.
During an internal evaluation by Tavus, 48 percent of participants who engaged in a one-minute video call with Griffin believed they were speaking to an actual human. Prior video avatar implementations reached a peak of just two percent on the same metric.
Benchmarks, Capabilities, and Rollout
Griffin processes both incoming and outgoing video streams concurrently. The model tracks and generates multiple communicative signals in real time, including tone of voice, spoken words, facial expressions, physical gestures, and natural conversational pauses.
In a test conducted by Nvidia evaluating human likeness during live audio-video chats, Griffin scored 3.83 points out of a scale where actual humans averaged 3.92. By comparison, the previous industry-best AI model reached 2.80 points.
Tavus has begun rolling out an early version, dubbed Griffin-Lite, to select testers as a research preview. The company plans to release a more capable iteration once it resolves associated safety concerns. Tavus is targeting several practical applications for the system, including interactive tutoring, role-playing difficult personal or workplace discussions, and visual tech support over video.
Evolution of Tavus
Founded in 2020, Tavus has raised roughly $64 million in venture funding. The company initially built automated video tools designed to generate personalized sales and marketing outreach at scale. It has since shifted focus toward interactive AI personas capable of carrying out live, bidirectional video dialogue.
