Tavus says Griffin fooled 48% of testers in one-minute video calls

Tavus reported that its Griffin-Lite model was mistaken for a real person by 26 of 54 participants (48%) in one-minute live video calls, a study the company calls the first pass of a real-time video Turing test. The October 1 announcement on X introduces Griffin-Lite as a research preview limited to selected testers, not a product available to Tavus customers, while Tavus develops disclosure and safety measures. In the same protocol, 1 of 41 participants (2.4%) mistook Tavus's previous Phoenix-4.5 system for a person, and NVIDIA ranked Griffin-Lite first on both the generation and perception tracks of its Video Full-Duplex Benchmark in September, scoring 3.83 and 3.73 out of 5 respectively.

Tavus says Griffin fooled 48% of testers in one-minute video calls The 26 of 54 participants who mistook Griffin-Lite for a person took part in Tavus's live study; the research preview remains limited to selected testers. By Ryan Merket https://runtimewire.com/author/ryan-merket · Published Primary source: Tavus on X https://x.com/tavus/status/2105704169009246248 Why it matters Griffin's study gives Tavus a striking result for its shift from generated avatar videos to live, visually aware AI conversations. But the sample was only 54 people, and Griffin remains unavailable to customers while Tavus develops disclosure and safety measures. Tavus says its Griffin model was mistaken for a real person by 26 of 54 participants in one-minute live video calls, a result the company calls the first pass of a real-time video Turing test. The October 1st announcement on X https://x.com/tavus/status/2105704169009246248 introduces Griffin-Lite as a research preview for selected testers, not a product available to Tavus customers. https://x.com/tavus/status/2105704169009246248 https://x.com/tavus/status/2105704169009246248 For Hassaan Raza @hassaanraza97 https://x.com/hassaanraza97 , Tavus's co-founder and CEO, Griffin extends a thesis he has put forward for years: computers should adapt to face-to-face human communication instead of making people learn a machine's commands. Raza's company vision essay https://www.tavus.io/vision traces that idea from command lines to graphical interfaces and argues for computers that can understand people through natural interaction. Tavus, which Raza co-founded with Quinn Favret, joined Y Combinator's Summer 2021 batch. The reported 48% is a result from a small company-run study, not a broad measure of how often people would mistake Griffin for a human in ordinary use. Tavus says participants in the United States and Europe were told they would speak with another participant for a minute about what they were looking forward to that year. They were asked whether their partner was real only at the end. The partner was a video persona powered by Griffin-Lite. In a comparison using the same protocol, Tavus says one of 41 participants, or 2.4%, mistook its previous Phoenix-4.5 system for a person. That study tests human perception in a live call. A separate evaluation tests system performance on NVIDIA's Video Full-Duplex Benchmark, which scores how models interpret and produce audio-visual conversational behavior. Tavus says NVIDIA scored Griffin-Lite in September and ranked it first on both the generation and perception tracks. On generation, Griffin-Lite scored 3.83 out of 5, against 2.80 for the next-highest system and 3.92 for the human reference. On perception, it scored 3.73, compared with 3.44 for the strongest reported baseline and 4.20 for the human reference. The scores came from a language-model judge; the benchmark result and the live study measure different things. Griffin's product bet is about the timing and visual context of conversation, rather than a face that simply speaks generated lines. Tavus describes a system that continuously processes incoming audio and video, decides whether to speak, wait or yield, and generates voice and video together. In company demonstrations, Griffin plays Simon Says, coaches a person solving a Rubik's Cube and responds to an object held up during a conversation. Tavus says the model generates full scenes rather than only animating a face. It also reports that the video-generation component produces a frame response to incoming audio in an average of 0.43 seconds on NVIDIA H100 GPUs, about half the latency of the next-fastest published system it compared against. The company is holding the preview back from customers because the same realism that makes Griffin useful can also make it deceptive. Tavus says it is working on disclosure features and further safety procedures before a wider release. That leaves a practical gap between the result and a deployable service: the study shows that some participants mistook Griffin for a person, while the company has not yet opened the model to general customer use. The timing fits Tavus's move from video generation toward live AI interaction. Tavus raised a $40 million Series B in November 2025, led by CRV with participation from Scale Venture Partners, Sequoia Capital, Y Combinator, HubSpot Ventures and Flex Capital, according to the company's funding information https://www.tavus.io/lp/ai-info-page . Tavus's existing platform plans https://www.tavus.io/pricing include paid access to conversational video, but Griffin itself is not currently offered through those plans. Griffin therefore serves first as a research and product direction for Tavus's real-time interaction business, with access and safety still unresolved before customer deployment.