What Griffin does
Griffin processes speech, facial expression, intonation, gestures and pauses. It takes in and generates video at the same time, allowing it to hold a real-time conversation.
Tavus also cites an independent Nvidia test of how human AI seems in live audio and video dialogue. Griffin scored 3.83 out of 3.92 for humans. The previous best score for an AI model was 2.80; Tavus says earlier systems reached a maximum of 2% in tests where participants mistook them for people.
A research launch, not a broad release
Griffin-Lite is available to selected participants in a research launch. Tavus says a more advanced version will follow once it has resolved safety issues.
The company points to three possible uses:
The test is not the product
Tavus, a San Francisco startup founded in 2020, has raised about $64 million. It began with personalized AI videos for sales and marketing before moving into real-time video conversations with digital characters. The company described Griffin in a research report.
I think the striking result is also a narrow one: fooling nearly half of participants for a one-minute call says little about how Griffin performs across longer or more demanding interactions. Tavus has not provided those details here. For now, the gap between a strong perception score and a release still gated on safety is the story: the next test is not whether Griffin can seem human briefly, but whether it can be trusted in the situations Tavus wants it to handle.
Source: the-decoder.com
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X