An AI Avatar Just Fooled Nearly Half the People Who Talked to It

Tavus says its new Griffin model convinced 26 of 54 people they were talking to a real human during a one-minute video call, up from 2.4% for its previous model.

Get a summary in:
Laptop screen showing a video call with coffee cup nearby
Follow Teck Hustlers on Google
Add Teck Hustlers as a preferred source to see more of our stories in Google Top Stories.
Add as preferred source on Google
AA

Admin Alex
Writer
MA

M.Ali
Fact-Checker

Published: October 3, 2026 · Last updated: October 3, 2026

TL;DR: Tavus says its new Griffin model convinced 26 of 54 people they were talking to a real human during a one-minute video call, up from just 2.4% for its previous model under the same test.

Person using a tablet for a video chat

Fifty-four people got on a one-minute video call. Twenty-six of them walked away convinced they’d been talking to a real person.

That’s the actual sample size behind the headline number Tavus is using to introduce Griffin, which the AI video startup is calling the first “Human Interaction Model” to pass what it’s branding as a video Turing test. The company ran the study itself: 54 participants, one-minute live face-to-face calls, and a post-call question asking whether they believed the person on the other end was real. 48% said yes. Tavus’s previous model, Phoenix-4.5, scored just 2.4% under the identical protocol, which is the comparison doing most of the work in making 48% sound as dramatic as it does.

Griffin’s technical pitch is a genuine departure from how most conversational AI video has worked until now. The usual setup stitches together separate systems: a speech-to-text layer transcribes what you say, a language model generates a response, and a separate voice and video system renders the output. Each handoff between those systems adds latency and loses context. Griffin runs as a single, full-duplex video-to-video pipeline instead, meaning it sees, hears, interprets, and responds simultaneously, with its perception staying active even while it’s mid-sentence. A sudden visual change, a pause, or an interruption can alter what it says and does in real time, the same way a real conversation partner would react.

Two components sit underneath that pipeline, according to Tavus: a continuous conversational modeling engine that assesses sub-second inputs to decide when to interject or back-channel with a small “mm-hm,” and a unified audio-visual generation engine that renders speech, movement, and scene-level pixel adjustments together instead of layering them separately. From a single reference image, Griffin can apparently render an entire dynamic scene in real time, controlling gestures, shadows, and background shifts as the conversation moves.

AI commentator Emad Mostaque reacted to the announcement with a one-line verdict: “remote work is cooked.”

Not every reaction was that breezy. Tavus’s own announcement thread drew pushback, including, by the company’s own account of the reaction, people arguing that AI this convincing should be illegal. That reaction is worth taking seriously even if Tavus itself hasn’t addressed it directly. A 48% success rate in fooling someone during a one-minute call is a very different risk profile once you imagine it applied to a scam call, a fake job interview, or a fabricated video reference check, rather than a company demo.

Tavus has not pretended this number is independently verified. The company ran the study itself, the sample size is small at 54 people, and the early version available to developers right now, called Griffin-Lite, is limited to a select group of testers while Tavus says it builds out disclosure and safety features before a wider rollout. That’s a meaningfully different claim than an independent lab confirming people can’t tell, and the gap between those two claims matters a lot if 48% ends up being the number people repeat without the asterisk attached.

Pricing and a broader release timeline haven’t been announced. What has been demonstrated, even accounting for the small sample and the self-reported methodology, is that the jump from 2.4% to 48% under an identical test represents a real architectural improvement, not just marketing math. Whether that improvement gets deployed responsibly is a separate question from whether it’s technically real, and right now only the second question has a clear answer.

Related: Supabase Just Bought Turso to Feed Its AI Agent Database Boom and ChatGPT Will Now Dress You Before You Buy Anything.

Bottom Line: A 20-point jump from one model generation to the next is the kind of number that should make people nervous, not impressed, and the fact that Tavus is marketing it as a selling point rather than flagging it as a risk says something about where this industry’s incentives still point.

Follow Teck Hustlers: Facebook · X · Instagram · LinkedIn · TikTok · Pinterest

Follow Teck Hustlers on Google
Add Teck Hustlers as a preferred source to see more of our stories in Google Top Stories.
Add as preferred source on Google