Tavus says 26 of 54 people in one-minute video calls with its new Griffin-Lite model believed they had spoken to a real person. That is a striking result from a study Tavus ran, but it measures responses to a particular call and question. Griffin-Lite is still a research preview for selected testers, and Tavus says it is developing disclosure safeguards before customers can use it.

The big change

  • What changed: Tavus has moved its video assistant research toward a system that listens, watches and generates a face and voice during the same conversation. In its short, company-run calls, 26 of 54 people believed the model was a person.
  • Why it matters: For teams designing video agents, conversational timing and expressive video can make an interface easier to engage with. The same study shows why telling callers they are meeting AI cannot be left as an afterthought.
  • What to watch: Griffin-Lite is limited to selected research testers. Tavus says disclosure features and further safety procedures are in development; their form and effectiveness, along with customer access and Griffin-specific pricing, remain unannounced.

What the 26-of-54 result measured

In Tavus's October 1 research account, participants were told they would be matched with another participant for a one-minute video call about what they were looking forward to this year. Their partner was instead a Tavus AI persona powered by Griffin-Lite. After the call, people rated the exchange and were asked whether they had wondered if their partner was not real. Tavus then disclosed that the partner was AI.

Of 54 Griffin-Lite participants, 26 said the partner was a real person, or about 48%. Tavus says it used the same protocol for its earlier Phoenix-4.5, Sparrow-2 and Raven-1 system; one of 41 participants in that group said their partner was a person. Those are different groups of callers, not a before-and-after result from the same people. The announcement does not establish how this response would change over longer calls, different topics or everyday use where people know an AI agent might appear.

Tavus calls the result a pass of a real-time video Turing test. That is the company's description of its one-minute study, not a universal certification for human-like conversation. RuntimeWire's independent launch report likewise treats the 48% figure as a small, company-run result and notes that the preview is still unavailable to customers. The study is evidence about participant judgments under its stated conditions; it is not an independently replicated deception rate.

A separate benchmark tests interaction quality

The NVIDIA VideoFDB leaderboard gives Griffin-Lite a 3.83 out of 5 on generation and 3.73 on perception. The benchmark uses clips from real video calls and a rubric-based language-model judge. Generation assesses the appropriateness of the model's voice, face and conversational behavior; perception assesses its responses to a person's audio and video. NVIDIA's table places Griffin-Lite above the listed systems on both tracks, while human-reference scores are 3.92 and 4.20 respectively.

Those scores measure a different task from asking callers whether they think a partner is human. They also have different comparison sets and conditions: the generation table includes cascaded speech-to-avatar systems, while the perception table includes systems evaluated with audio and video or, in separate rows, audio alone. The leaderboard supports a result on NVIDIA's published rubric. It does not validate Tavus's 48% call-study figure or show how Griffin would perform as a customer service, tutoring or healthcare product.

What Tavus says the system does

Tavus describes Griffin as a two-part design. A continuous conversational model takes in a caller's audio and video and issues controls for what to say, when to speak, and how to express it. Streaming speech and video generators turn those controls into a voice and moving scene. The company says the system can listen while speaking, yield when interrupted and react to visual context instead of waiting for each turn to finish. These are Tavus's technical descriptions and demonstrations; BIG CHANGE has not accessed the preview or measured its behavior.

That distinction matters because a generated face alone is not the whole interaction. A system that responds to pauses, gaze and interruptions could change whether a caller feels understood or realizes they are speaking with software. The VideoFDB result offers a scored comparison of conversational behavior under a benchmark protocol. Tavus's human-call study offers a narrower observation about identity judgments. Neither substitutes for public evidence about sustained reliability, disclosure in a live deployment or users' informed consent.

Access and the decision ahead

Tavus says Griffin-Lite is open only to select trusted testers as a research preview and is not available to Tavus customers. The company offers a request form for testing. Griffin is not yet on its general platform or public API, and the announcement gives no Griffin-specific price. Tavus's existing Phoenix, Raven and Sparrow models remain a separate current product path.

The company says it is working on disclosure features and further safety procedures before a wider release. It has not published a date, detailed disclosure design or evidence that such measures work. For anyone evaluating a real-time video agent, the immediate question is therefore specific: how will a caller know who or what is on the other side, and how will that be checked before Griffin becomes a customer product?

Sources & further reading