Tavus Griffin: 48% thought the AI video agent was human

Tavus launched Griffin, a real-time AI video agent, on October 1, and the number in the headline is the one that matters: in a blind study, 48% of callers believed they were talking to a real person. That is close to a coin flip. For any business that wants a face on its support or sales bot, the hard question is no longer "will it look fake?" It is "will we tell people it isn't real?"

What actually happened

Tavus calls Griffin a "Human Interaction Model." Most video avatars today are a pipeline: speech-to-text, then a language model, then text-to-speech, then an animated face. Griffin is one full-duplex, video-to-video model. It listens, watches and talks at the same time, so it can nod or say "mm-hm" while you speak and stop when you interrupt.

The specifics, per Tavus's launch page and its Business Wire release:

• Turing-style study: 26 of 54 participants (48%) said they spoke with a person after a one-minute call. Tavus's previous system fooled 1 of 41.
• Benchmark: first on NVIDIA's VideoFDB full-duplex benchmark, with a generation score of 3.83 out of 5 against a human reference of 3.92.
• Output: 720p video generated in real time in 320 ms chunks from a single reference image, with about 0.43 seconds average audio-to-video latency.
• Access: Griffin-Lite is a research preview for "select trusted testers" only. Tavus says wider release waits on more disclosure and safety mechanisms. No API, no pricing yet.

Tavus says 150,000 developers and businesses use its platform and that it has raised $70M.

A 54-person, one-minute study is small, and Tavus ran it. Treat 48% as a direction, not a law.

Why it matters for your business

Realism is now the default, so disclosure is your job. When half your callers can't tell, "they'll figure it out" stops being a plan. Put "You're talking with an AI assistant" in the first sentence and on screen. California already requires bots to disclose themselves when they're used to drive a sale. Customers who feel tricked don't come back.

A face does not fix a bad answer. Griffin makes the conversation feel human. It does not make your return policy, inventory data or escalation path any better. Get the voice or chat agent right first. Add video when the content is solid.

Wait for the price. No API and no pricing means you can't budget it. Real-time 720p video on H100s will not cost what a text bot costs. Pilot only where a face changes the outcome, such as high-ticket consults or onboarding.

Sources: Tavus: Griffin, Tavus via Business Wire.