The Pulse
Tavus Says Griffin Passed Its Real-Time Video Turing Test
Tavus says 26 of 54 participants believed Griffin-Lite was a real person after a one-minute video call. The company is limiting access to select testers while it works on safety and disclosure features.
AI.info Team · Written with AI ·
Twenty-six of 54 people in a one-minute video call thought Tavus’s Griffin avatar was a real person, the company said October 1, 2026. Tavus calls the result a pass of the “real-time video Turing test,” a label for a short, face-to-face interaction in which participants did not know their partner was AI. The finding comes from a company-run study, not a general test of whether people can distinguish AI from humans across longer or different kinds of conversations.
What the 54-person test measured
Tavus says participants were recruited through an independent research platform. They were told they would speak by video with another participant for one minute about what they were looking forward to that year; instead, their partner was a real-time persona powered by Griffin-Lite. After the call, participants were asked whether it had crossed their mind that their partner might not be a real person and, if so, when. They were told their partner had been AI at the end of the survey.
Twenty-six participants answered that their partner was real, or 48% of the group. Tavus compares that result with a previous test of its Phoenix-4.5, Sparrow-2 and Raven-1 system: one of 41 participants, or 2.4%, believed that avatar was human. The company says the earlier test used the same protocol, though the two results come from separate participant groups.
The result captures how participants judged one brief interaction under a specific setup. It does not establish that Griffin would fool most people in longer calls, in other settings or when users know they may be speaking with an AI. Tavus says participants who suspected their partner was AI tended to do so within the first 20 seconds.
Griffin listens while it speaks
Tavus describes Griffin as a full-duplex video-to-video system: it processes incoming audio and video while generating its own speech and video, rather than waiting for a person to finish before responding. The company says the model makes decisions at sub-second intervals about whether to speak, wait or yield the floor, and can react to interruptions and visible behavior. Its output includes facial expression and gestures, not just spoken words.
The system combines a conversational model, which interprets the exchange and controls how Griffin responds, with streaming audio and video generation. Tavus says Griffin-Lite generates 720p video in 320-millisecond chunks. Those technical design claims describe how the system works; the one-minute participant study is the evidence the company offers for its headline claim about human-likeness.
A research preview, not a customer release
Tavus is making Griffin-Lite available only to select trusted testers and says it is not currently available for customer use. The company says it is developing safety and disclosure features and believes further alignment work is needed before a safe release. Griffin’s ability to be mistaken for a person is the reason Tavus gives for limiting access.
“Humans are evolutionarily designed to communicate face to face.”
Hassaan Raza, co-founder and CEO of Tavus
The announcement leaves two concrete questions: how Tavus will make AI disclosure clear during a live exchange, and whether Griffin’s result holds up in larger studies with longer calls. For now, the company’s published figure is 26 of 54 participants after a single minute.