Tavus’ Griffin AI convinced 26 of 54 people on live video calls that they had talked to a real human. The firm is keeping the model away from customers for that exact reason.
Tavus calls the model the first to pass the video Turing test. Its previous system, tested the same way, fooled 1 of 41 participants.
What Let Griffin Talk Its Way Past 26 Strangers?
Tavus co-founder and CEO Hassaan Raza argued that people communicate as much through expression, tone, and timing as through words. Machines, by his account, make users adapt with the right command, click, or prompt.
His pitch is a computer that reads tone and expression the way another person would, and even looks like one.
Tavus contrasts Griffin with cascade systems, where speech recognition, a language model, and voice synthesis take turns. Such systems hold their reply until the user has finished speaking.
Griffin re-checks the conversation every fraction of a second, reading words and nonverbal cues together. It watches the video too, which Tavus says most interactive systems skip.
That design lets it nod or say “mm-hm” while someone is still talking, and it keeps its place when interrupted.
“Griffin-Lite’s speech generation produces high-quality speech at conversational latency and can clone a speaker’s voice from about 10 seconds of audio,” Tavus claims.
NVIDIA’s VideoFDB benchmark tests how well AI holds a live video conversation. It scored Griffin 3.83 out of 5.
Human reference recordings scored 3.92, while the next-best system, a Gemini 2.5 and Anam combination, reached 2.80.
Why Is Tavus Keeping Its Griffin AI From Customers?
Study participants expected a one-minute chat with another person about their year ahead. More than half said the idea of an AI partner never crossed their minds.
Those who did grow suspicious usually did so within the first 20 seconds. Tavus acknowledges that the same realism could let the model deceive people into believing they are talking to a human.
Therefore, Griffin-Lite, the preview version, is open only to select trusted testers. The company says it is building disclosure features and working with AI safety organizations before a wider release.
Crypto fraud shows why that risk is not hypothetical. TRM Labs found criminal AI adoption rose 40% year-on-year in 2026, with scams leading that growth.
If models like Griffin spread, a familiar face on a video call may prove much less about who is calling.
Subscribe to our YouTube channel to watch leaders and journalists provide expert insights