October 1, 2026
Griffin-Lite mistaken for a human: 48% of Tavus study participants fooled after a video call
In a Tavus study dated Oct 1, 2026, 26 of 54 participants mistook Griffin-Lite for a human after a one-minute video call.
🚨 AI News | TestingCatalog
@testingcatalog
Tavus introduced Griffin, a research preview of its new video-to-video model for real-time face-to-face conversations. Griffin learns how people behave in conversation. The model can watch and listen to people, interrupt them mid-sentence or respond when someone enters the frame. > Tavus calls Griffin the first model to pass the Turing test: 45% of the study's 120 participants mistook a conversation with it for a conversation with a real person. > Griffin Lite ranks first among all models evaluated in the NVIDIA Video Full-Duplex Benchmark: 3.83 out of 5 versus 3.92 for real people. > The full Griffin model has yet to be released: Tavus is reconsidering its release policy.
Introducing Griffin, the first model to pass the video Turing test. 48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video. It’s the first Human Interaction Model (HIM).
· 11.3K views
Griffin holds video conversations: it watches, listens and can interrupt someone mid-sentence or react when a person enters the frame.
Tavus trains the model on human conversational behavior. In the demo, Griffin helps Sagar solder a motherboard: it tracks the stage of the work and the time, and suggests when to move on to the next step.
Before and now. In Tavus's study dated Oct 1, 1 of 41 participants mistook the previous Phoenix-4.5 + Sparrow-2 + Raven-1 combination for a human, or 2.4%. Griffin-Lite scored 48% after a one-minute conversation.
Griffin Lite also ranked first among models in both NVIDIA VideoFDB tracks as of Oct 1, 2026. One track evaluates video response generation, while the other evaluates perception of the conversation partner.
| NVIDIA VideoFDB track | Participant | Score out of 5 | | --- | --- | --- | | Generation | Griffin Lite | 3.83 | | Generation | Human reference | 3.92 | | Generation | Gemini 2.5 + Anam | 2.80 | | Perception | Griffin Lite | 3.73 | | Perception | Human reference | 4.20 | | Perception | MiniCPM-o 4.5, audio only | 3.44 |
According to Tavus's measurements on H100 dated Oct 1, Griffin-Lite reflects incoming audio in video after an average of 0.43 seconds. This is twice as fast as the next-fastest method in the company's comparison.
Griffin-Lite generates the entire frame, including the body and background, from a single reference image. According to Tavus's data dated Oct 1, the model generates 720p video at 25 frames per second in 320 ms chunks and clones a voice from an audio recording lasting roughly 10 seconds.
How to get access. Griffin-Lite is available to selected testers in a research preview. Tavus accepts applications through a form in the Safety, limitations, and responsibility section of its Oct 1 announcement. The model is not yet available on the Tavus platform.
Tavus has yet to release the full Griffin model: the company is reconsidering its release policy.