Artificial Intelligence

Study: Half of Testers Thought Tavus' AI Video Caller Was Human

Is this a record for AI-to-human conversation, a hyped up marketing stunt or a danger to society either way?

2 hours ago·Updated 1 hour ago

#Artificial Intelligence

Study: Half of Testers Thought Tavus' AI Video Caller Was Human

Story

Above: A stock photo illustration representing an AI-generated brain. Image credit: Zach Malinowitz/Unsplash

The Facts

Tavus, a San Francisco-based AI research company, unveiled its system called Griffin on Oct. 1, describing it as the first Human Interaction Model, a full-duplex video-to-video system that sees, hears, speaks and reacts during live face-to-face conversations.
Infobae
Business Wire
In a company study, 26 of 54 participants, or 48%, believed Griffin-Lite was a real person after a one-minute video call. Tavus' earlier system, combining Phoenix-4.5, Sparrow-2 and Raven-1, convinced 1 of 41 participants, or 2.4%.
The Rundown AI
Tavus
Participants were recruited through an independent research platform and told they would be matched with another participant to discuss what they were looking forward to this year. Only at the survey's end were they asked whether they suspected an AI.
Infobae
XenoSpectrum
On NVIDIA's VideoFDB benchmark, Griffin-Lite scored 3.83 out of 5 on the generation track, against 3.92 for human reference recordings and 2.80 for the next listed system, Gemini 2.5 paired with Anam.
The Rundown AI
Business Wire
On the perception track, which scores whether the system reads the other person's cues and answers appropriately, Griffin-Lite scored 3.73 out of 5, compared with 4.20 for humans. Its takeover-rate alignment — how often it spoke or waited at the moments the reference calls called for — was 73.8% versus 90%, and its median response delay was 2,232 milliseconds versus 1,400.
Tavus
XenoSpectrum
Tavus is limiting Griffin-Lite to select trusted testers as a research preview while it develops disclosure and safety features, and has not given a public launch date. The company says more than 150,000 developers and businesses use its earlier models.
Tavus on X
Business Wire
The Rundown AI

Sources Split

LEFT

CENTER

RIGHT

ESTAB
CRIT

CENTER

PRO
ESTAB

The Spin

Techno-optimist narrative

Whether this passed the Turing test perfectly or not, teaching a machine to read a pause, hold a glance and answer in the same beat a person would is the breakthrough that actually matters. Independent scoring by an outside lab put this system within a hair of human performance and a full point ahead of everything else. Fooling half a room on a live call is the clearest proof yet that face-to-face machine conversation has arrived.

SlopTV News

Tavus

Techno-skeptic narrative

A one-minute chat with 54 people recruited and graded by the company selling the product is not at all close to a Turing-level performance. Nobody was told to look for a machine; there was no human control group and the suspicion rate spiked in the first 20 seconds. On the measures that count — timing, knowing when to speak and nonverbal fit — the gap to an actual person is still wide.

CellCog

XenoSpectrum

Cynical narrative

There's no good reason to create AI that can fool people into thinking they're talking to a real human. Similar chatbots — which, according to this latest release are less advanced — have already tricked an elderly man into traveling to visit a fake woman and a teenager to commit suicide. With that already happening, the future consequences are endless, from scamming people into providing bank information to kids growing up with AI boyfriends and girlfriends. This entire field of AI should be put to an end.