MEMES
IS NEWS

The news is serious. We don't have to be.

MEME #00060

AI just video-calls you back: 48% thought it was a real human

Tavus announced Griffin on Oct 1, 2026, the first AI model that holds real-time, face-to-face video calls. In Tavus's own study, 26 of 54 people (48%) on a one-minute video call b...

Published October 3, 2026 · AI

#LILMEME00060 #MemesISNews #Tavus #AI #TuringTest

Read the original caption

Tavus announced Griffin on Oct 1, 2026, the first AI model that holds real-time, face-to-face video calls. In Tavus's own study, 26 of 54 people (48%) on a one-minute video call believed they were talking to a real human - up from 2.4% for its previous system. Independent NVIDIA VideoFDB scoring put Griffin at 3.83/5 for generation quality against 3.92 for a real person. Because it is the first model people mistake for a person, Tavus will not release it yet and says more disclosure and safety work is needed. Senator Bernie Sanders called it 'insane'.

The real news

The story, straight

Tavus announced Griffin on October 1, 2026: the first AI model that holds real-time, face-to-face video calls. In Tavus's own study, 26 of 54 people (48%) on a one-minute video call believed they were talking to a real human - up from 2.4% for its previous system.

Independent NVIDIA VideoFDB scoring put Griffin at 3.83 out of 5 for generation quality, against 3.92 for a real person. Tavus says it will not release it yet because it is the first model people mistake for a person, and that more disclosure and safety work is needed.

Senator Bernie Sanders called it 'insane.'

The absurd

Why this is ridiculous

48% of people looked at this face on a video call and believed it was a person. The previous model fooled 2.4% of them.

That is the distance between 'that's clearly a robot' and 'that might be a person' - and it is one release away.

The breakdown

How it breaks down

The 2.4% to 48% jump is the number that matters, not the 3.83. It is not a small smoothness gain; it is a category change.

Once a model can hold a face on a live call, detection stops being a skill people have - so disclosure becomes the entire safety mechanism. Tell them, or they can never know.

The upside

Why it could matter

The upside is the same math aimed at a person: real-time, face-to-face AI can be a companion, an interpreter, a presence for someone who is alone. The rule that keeps it a gift instead of a trap is the one Tavus actually adopted - say it is AI.

The trick only works when everyone plays by that line.

Sources

Tavus, Oct 1 2026 (Griffin announcement) + ThePrint, Oct 3 2026 (48% of 54 testers on a one-minute video call believed it was human; 2.4% for prior Phoenix-4.5) + NVIDIA VideoFDB benchmark (3.83 vs 3.92 human)

← Back to the archive