Tavus introduced Griffin on October 1 as a model that treats a live video conversation as a continuous exchange. Its announcement combines speech, facial behavior, gestures and interruption handling in a video-to-video system, with Griffin-Lite in preview. The headline result is about how human the interaction appeared during a short study. Tavus’s announcement
Tavus reports that 26 of 54 participants, or 48%, believed their Griffin-Lite conversation partner was a real person. In a comparison using its earlier Phoenix-4.5, Sparrow-2 and Raven-1 system, one of 41 participants, or 2.4%, did. The company describes Griffin as passing a real-time video Turing test. That label is Tavus’s interpretation of the reported study, rather than an independently reproduced certification.
The protocol is important. Participants were recruited through an independent research platform and told they would speak with another participant. The call lasted one minute and concerned what they were looking forward to that year. They were asked about their partner’s humanity afterward, and all were subsequently told the partner was AI. This establishes a specific finding about impressions in that setting; it does not show that users would form the same impression when told in advance that they were talking to a model.
The product’s underlying bet is that listening and speaking cannot be reduced to alternating text turns. Tavus describes continuous perception and generation that can coordinate a nod, pause or interruption with the conversation. Such timing could make an interaction feel more responsive even when the factual content is unchanged.
My inference is that conversational naturalness and practical competence need separate evaluation. A convincing face and voice can help an interaction flow, but the one-minute result does not establish accuracy, reliability over longer calls or successful completion of demanding tasks. It also does not establish that perceived trustworthiness is warranted.
Griffin’s announcement provides an unusually concrete study size and comparison for its naturalness claim. The next product question is whether that improvement helps people accomplish their intended work while clearly understanding whom they are speaking with. The reported impression is meaningful within its protocol, and that protocol should travel with the headline number.