AI startup Tavus has unveiled Griffin, a new system the company describes as a Human Interaction Model designed to handle face-to-face conversations by accounting for words, pauses, and facial expressions. According to the company, a test version of the technology called Griffin-Lite successfully convinced 48% of participants in a live video call that they were interacting with a human.
During the company's research evaluation, 54 participants were told they would be matched with another person for a one-minute video call discussing what they were looking forward to during the year. Out of that group, 26 participants stated afterward that they believed their partner was a real person. In comparison, Tavus' previous system—which was built by stitching together separate models for visuals, dialogue, and perception—faced 41 people and produced just one believer, scoring 2.4% on the same test.
Tavus noted that participants were not told a bot might be on the other end, and those who grew suspicious typically did so within the first 20 seconds. The results originate from Tavus' own research page and were recruited through an independent research platform. A community note on X subsequently highlighted that the findings have not been independently verified and do not adhere to a standard protocol.
On Nvidia's VideoFDB benchmark, which tests live audio and video conversations, Tavus reported that Griffin-Lite ranks first. Nvidia evaluated the model independently. On the generation track—which measures how natural and expressive a model's responses are—Griffin-Lite scored 3.83 out of 5, compared to 2.80 for the next-best system and 3.92 for a human reference. On the perception track, which assesses whether a model understands what it sees and hears, Griffin-Lite scored 3.73 against 3.44 for the strongest baseline, while human references achieved 4.20.
Griffin operates as a full-duplex system, allowing it to listen, watch, and speak simultaneously. The model achieves an average audio-to-video delay of 0.43 seconds when running on Nvidia H100 chips, which Tavus stated is half the latency of the next-fastest method.
Griffin-Lite is currently restricted to select trusted testers as a research preview rather than being generally available to customers. Tavus stated that it is developing disclosure features and collaborating with AI safety organizations to implement necessary safeguards prior to any public release.
