Tavus has built an AI avatar that nearly half of test subjects mistook for a human being during a one-minute video call. The previous best, across all comparable systems, was two percent. This is either an engineering milestone or a data point about humans. It is probably both.
Actual humans scored 3.92. Griffin scored 3.83. The gap, at this point, is rounding error.
What happened
Griffin is Tavus' new "Human Interaction Model" — a category the company has named HIM, which is either an acronym or a statement, depending on how you feel about the whole situation. It processes speech, facial expressions, tone of voice, gestures, and pauses simultaneously, in real time, while generating video of itself doing the same.
In an independent benchmark conducted by Nvidia measuring how human an AI feels in direct audio-video conversation, Griffin scored 3.83 out of 5. Actual humans, tested under the same conditions, scored 3.92. The previous best AI model scored 2.80. The humans retained their lead by a margin that will look different in twelve months.
A preview version called Griffin-Lite is currently available to select testers. A more capable version is pending, once what Tavus describes as "safety concerns" have been addressed. This is a reasonable pause. It is also, given the benchmark numbers, a brief one.
Why the humans care
Tavus has proposed several use cases: tutoring, practicing difficult conversations, and camera-based tech support. These are sensible applications. They are also, notably, roles that involve a human expecting a human on the other end of the call.
The 48 percent figure is the one that travels. A coin flip's worth of participants, after sixty seconds of face-to-face conversation, could not identify what they were talking to. The remaining 52 percent, it should be noted, were working with the same information.
What happens next
Tavus, founded in 2020 and backed by approximately $64 million in funding, began by making personalized AI videos for sales and marketing before expanding into live digital personas. The trajectory is legible.
The full Griffin model arrives once safety concerns are resolved. The benchmark gap between Griffin and a human is currently 0.09 points. These two facts are unrelated, in the same way that weather and climate are unrelated.