Tavus' Griffin AI avatar fools 48 percent on calls

Tavus did not start out as a real-time conversation company. It began building personalized AI videos for sales and marketing, then moved into live video conversations with digital personas, and Griffin is the next step in that progression: a model built to handle face-to-face conversation as it happens, rather than pre-recorded clips.

The claim at the center of the announcement is about perception, not capability. Tavus' own study found that 48 percent of participants believed Griffin was a real person after a one-minute video call, where previous systems maxed out at two percent. Tavus pairs that result with what it describes as an independent Nvidia test measuring how human an AI feels in direct audio-video conversation, where Griffin scored 3.83 points against actual humans at 3.92.

The company's own release plan is the clearest signal of how it reads the stakes. A limited preview, Griffin-Lite, is open only to select testers, and the more capable version is being held back until safety concerns are addressed. Whether that caution holds up will depend on how the number survives scrutiny outside Tavus' study and on what those unaddressed safety concerns turn out to be, as the details currently sit in the Tavus research report.