Tavus says 48% of 54 participants mistook its Griffin AI video avatar for a real person during a one-minute video call. That is a striking result for a system designed to react in real time, and a reminder that disclosure cannot be an afterthought when synthetic faces become convincing. The Decoder reported the findings and product details.
Watch Desk analysis
What happened
Tavus has introduced Griffin, which it describes as a real-time model able to watch, listen and respond during face-to-face video conversations. The company says Griffin ranked first on NVIDIA’s full-duplex video benchmark and claims it is the first model to pass a real-time video Turing test.
The 48% result comes from a company study of 54 people. A preview called Griffin-Lite is available to select testers, while Tavus says it is withholding a general release as it develops safety procedures and disclosure features. The Decoder’s report on Griffin and the study carries the details.
Why it matters
A convincing face and voice in a live call can make an AI system feel less like a tool and more like another person in the room. That may be useful for interactive services, but it also raises a plain question: will people know when they are speaking to a generated avatar?
The study is a small company-run test, not proof that most people would be fooled in everyday use. Even so, its reported result gives the disclosure question a concrete edge. Tavus’s decision to delay general availability while it works on safeguards suggests the company sees that issue too.
Our read
The interesting part is not a Turing-test trophy. It is that nearly half of a small test group reportedly missed the distinction in a one-minute call. That makes clear, persistent disclosure part of the product, not a decorative label tucked out of sight. For now, Griffin-Lite is limited to select testers, so this is a signal to watch rather than a tool most people can try today.
What to watch
- What disclosure features Tavus develops before a wider launch.
- Whether Griffin-Lite remains limited to selected testers or expands.
- How the company describes the study’s method and its limits.
Discussion spark: If an AI avatar can pass as a person in a short call, should the platform be required to keep an unmistakable disclosure visible throughout the conversation?
Sources and evidence
- Nearly half of test subjects mistook Tavus' AI video avatar for a real person on a one-minute call (1 October 2026, 18:03 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.
