Tavus says 26 of 54 people took its Griffin-Lite AI for a real person on a one-minute video call

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude. Google and OpenAI, whose realtime models appear in the benchmark table discussed below, are Anthropic competitors.

Tavus, a San Francisco company that makes face-to-face AI characters it calls PALs, published a research post dated 1 October 2026 introducing Griffin-Lite, which it calls a preview of its "first Human Interaction Model". Tavus says that in its own study, 26 of 54 participants (48%) who had a one-minute live video call with the model believed they had been talking to a real person.

What Tavus says it built

Tavus describes Griffin-Lite as "a full-duplex video-to-video model that responds to human behavior in real time". In its words, perception, "deciding when and how to respond, and expressive speech and video generation all happen at the same time".

Its technical claims, all Tavus's own:

• It makes conversational decisions "at regular sub-second intervals rather than once per turn".
• It "can clone a speaker's voice from about 10 seconds of audio".
• It "generates 720p video in 320 ms chunks in real time" from "one reference image".

The study, in Tavus's words

Tavus says participants were "recruited through an independent research platform" (unnamed) and "told they would be matched with another participant for a one-minute video call to discuss what they were looking forward to this year". Their partner was an AI character running on Griffin-Lite. Only at the end of a survey were they asked "whether it had crossed their mind that their partner might not be a real person", and all were then told it had been an AI.

The results Tavus reports:

• Griffin-Lite: 26 of 54 said their partner was a real person (48%).
• Earlier system (Phoenix-4.5 with its Sparrow-2 and Raven-1 models), same protocol: 1 of 41 (2.4%).
• "Over half of participants said the possibility had not crossed their mind during the call". Those who did suspect "tended to suspect within the first 20 seconds".

Tavus calls this, "to the best of our knowledge, the first model to have ever passed the video Turing test". Elsewhere on the same page it drops the qualifier and says Griffin "is the first model to pass the real-time, video Turing test". Both are Tavus's claims about a test it designed and ran.

How much weight the number can bear

This is a vendor-run study of 54 people on the Griffin-Lite side. The page does not cite a peer-reviewed paper or publish the raw responses.

Our arithmetic, not Tavus's: a 95% Wilson confidence interval for 26 out of 54 runs from about 35% to 61%. The "48%" is consistent with anything from roughly a third to three-fifths.

On our reading, the set-up also matters. Participants expected another person and were asked about AI only afterwards; in the three-party version used in a 2025 study (arXiv 2503.23674), "Participants had 5 minute conversations simultaneously with another human participant and one of these systems before judging which conversational partner they thought was human." Tavus's set-up is a different, arguably easier bar, over one minute.

The NVIDIA benchmark

Tavus also says Griffin-Lite is "#1 on NVIDIA's independent test of face-to-face AI": VideoFDB, a benchmark from NVIDIA and David AI. Tavus says "NVIDIA conducts the evaluation independently using the published metrics and its own judge", and that NVIDIA scored the results in September 2026.

NVIDIA's own VideoFDB leaderboard does list "Tavus Griffin Lite" first among the models in both tables (Tavus's own chart puts Gemini 2.5 Flash Native at 3.17 and OpenAI gpt-realtime at 2.97 on perception; that is gpt-realtime's audio-only run, and NVIDIA's table gives it 2.75 with video):

• Perception table (15 model entries plus a human reference): Griffin-Lite 3.73 out of 5, against 3.44 for the next entry (MiniCPM-o 4.5, audio-only) and 4.20 for the human reference.
• Generation table (three model entries plus human ground truth): Griffin-Lite 3.83, against 2.80 for Gemini 2.5 + Anam and 2.39 for Gemini 2.5 + Keyframe; human ground truth is 3.92.

The generation lead is over two cascaded systems, the only other entries in that table. Tavus's headline "37% ahead of the next best AI" matches, on our arithmetic, the ratio of 3.83 to 2.80. Both pages say a language-model judge produced the scores. Neither describes the listing as an NVIDIA endorsement.

Availability and safety

Tavus says Griffin-Lite "will not be available for use for customers at this time", only to "select trusted testers as a research preview". It writes that such models can "deceive a human into believing it is not AI", says it is "working on safe disclosure features", and anticipates release "very soon after these safety concerns are addressed".

Checking a video call yourself

UK police guidance already treats video as fakeable. A Surrey Police leaflet on AI and deepfakes, published by the county's Police and Crime Commissioner (information current as of November 2025), warns that scammers "can create hyper-realistic video of 'family members' or 'friends' in distress" and advises: "Pause, think and verify the communication independently through a trusted separate channel". It also suggests calling "them back directly on their known number", or setting up "a family code word".

The government's Stop! Think Fraud campaign gives similar advice for calls: agree "a safe phrase" with close friends and family, and "don't trust the Caller ID display on your phone – it's not proof of ID". The Surrey leaflet lists Action Fraud on 0300 123 2040; City of London Police says Report Fraud replaced Action Fraud as the national reporting service from 4 December 2025, and it uses the same number for people who live in England, Wales or Northern Ireland.

Why it matters

On our reading, the 48% figure rests on a small vendor-run sample, but the direction is clear from Tavus's own account: a model it says works from one photo and a short voice clip, which it is holding back from customers over deception risk. The UK guidance above, which predates this release, already treats a face on a screen as no proof of who is speaking.