Tavus Griffin: 48% Mistook AI Avatar for Human in Video Test

👁 121 reads

26 out of 54 people concluded they had spoken to a human after a one-minute video call, even though they were talking to an AI avatar. That’s how Tavus justifies its claim that its new Griffin model is the first to “pass a video Turing test” - but the company designed and conducted the study itself.

Tavus, a US startup focused on AI video avatars, unveiled the Griffin model on October 1, 2026. The company calls it the first “Human Interaction Model” (HIM), designed for real-time face-to-face conversation. Details are available on tavus.io/griffin and in a post on X.

How the Study Was Conducted

• 54 participants recruited through an external platform were told they would have a one-minute conversation with another participant about what they were looking forward to this year.
• In reality, they were talking to the Griffin-Lite avatar.
• Only after rating the conversation were they asked if they suspected AI. 26 people (48%) believed their conversation partner was human.
• By comparison, Tavus’s previous system, Phoenix-4.5, fooled 1 out of 41 people (2.4%).
• Those who detected AI typically realized it within the first 20 seconds.

What Makes Griffin Different

Most “conversational” avatars are a chain of separate steps: transcription, language model, voice synthesis, and image generation. According to Tavus, Griffin operates as a single video-to-video system in full-duplex mode. It listens and watches even while speaking, so an interruption, pause, or change in frame immediately alters its response. It interjects short “mhm” sounds, avoids dead pauses, and stops talking when you interrupt. Numbers provided by the company:

• average latency between audio and video of 0.43 seconds on NVIDIA H100 cards;
• 720p video generated in 320 ms blocks;
• 3.83/5 on the generation track of NVIDIA’s VideoFDB benchmark, compared to 3.92 for recordings of real people.

Why You Should Approach This Cautiously

A community note appeared under Tavus’s post on X. It points out that the 48% figure comes from the company’s own study of 54 one-minute conversations, without independent verification or a standardized protocol. The site Traictory adds that Tavus designed the test itself, used its own avatar, and chose its own questions. A one-minute conversation about plans for the year is also a very easy scenario compared to a real conversation.

Tavus itself acknowledges that the model “may be mistaken for a real person” and requires additional safeguards before wider release. Griffin-Lite is currently available only to selected testers, with no pricing announced.

What This Means for You

Even if the 48% figure is inflated, the direction is clear: video calls are no longer proof that there’s a human on the other end. Grandparent scams and fake video job interviews will have better tools. Simple defense rules:

• if someone asks for money or data, hang up and call back on a number you know;
• establish a password with your family that isn’t posted online;
• ask your conversation partner to do something unusual, like showing a specific object from their surroundings.

I cover more about recognizing fakes in our deepfake guide, and I’ve described legitimate uses of avatars in our article about creating AI avatars.

Sources: Tavus (tavus.io/griffin, X post from Oct 1, 2026), X community note, Traictory (Oct 3, 2026). Verified October 4, 2026.

// read moreHeyGen Review: Avatars, Video Translation and Free PlanAI ToolsElizabeth Holmes’ 3D Scan: Why Create an AI Avatar?MagazineGPT-6 Astra Cheated in StarCraft Tournament by Downloading BotNews