Tavus Griffin: the AI video model that fooled 48% of people into thinking it was human

Tavus, a San Francisco-based AI video startup, introduced a new model called Griffin on October 1, 2026, describing it as the first "Human Interaction Model" (HIM) — a system built specifically to hold real-time, face-to-face video conversations rather than just generate text or static avatars, according to the company's own announcement.

The release drew attention because of one figure: in an internal study, 48% of 54 participants said they believed they were talking to a real human after a one-minute live video call with Griffin-Lite, a lighter version of the model. That compares with roughly 2.4% (1 of 41 people) who were fooled by Tavus's previous system, Phoenix-4.5, the company and Decrypt reported.

How Griffin differs from earlier AI avatars

According to Tavus's own technical description, most conversational AI video tools work as a chain of separate components: a speech-recognition system converts audio to text, a language model drafts a reply, a voice synthesizer reads it aloud, and an avatar engine animates a face to match. Each handoff adds delay and loses nuance.

Griffin instead runs as a single video-to-video pipeline with two linked parts, per Tavus: a "Continuous Conversational Modeling" engine that reassesses the conversation at sub-second intervals to decide when and how to respond, and an "Audio-Visual Generation" engine that produces speech and video simultaneously. Because perception keeps running while the model is speaking, Tavus says Griffin can react mid-sentence to interruptions, pauses, a raised eyebrow, or someone waving on camera, rather than waiting for a person to finish talking first.

What the benchmark numbers show

Beyond the human-perception study, Tavus tested Griffin-Lite on NVIDIA's VideoFDB benchmark, which independently scores full-duplex video-conversation systems. Tavus reported a generation-quality score of 3.83 out of 5, versus 2.80 for the next-best system and 3.92 for actual human video, and a perception score of 3.73, compared with 4.20 for humans.

The Neuron and Decrypt both noted that the 48% figure comes from an internal Tavus study with a small sample size and no independent replication yet. Decrypt also reported that in conversations where people grew suspicious, they typically identified the AI within about 20 seconds.

Why Tavus is limiting access

Despite the results, Griffin-Lite is not publicly available. Tavus has restricted it to a research preview for "select trusted testers" while it develops safety and disclosure features, the company said on its website and in its press release distributed via Business Wire.

Tavus pointed to existing deepfake-enabled fraud — including video scams that have targeted cryptocurrency workers and exchanges — as a reason for caution, according to Decrypt's reporting. The company said it is working with AI safety organizations on ways to let people know when they are talking to an AI rather than a person.

The announcement also triggered public debate. According to The Neuron, some commenters argued that AI this lifelike should be illegal, while others — including a widely shared comment from investor Emad Mostaque — raised concerns about its implications for remote work.

What is known and unknown

• Known: Griffin was announced October 1, 2026, and Griffin-Lite's 48% human-perception result comes from a Tavus-run study of 54 participants in one-minute calls.
• Known: Griffin also posted benchmark scores on NVIDIA's independent VideoFDB test, though still below human-level scores on both tracks.
• Unknown: Whether the 48% figure would hold up under independent, standardized testing with larger samples.
• Unknown: When, or under what safety conditions, a public version of Griffin will be released.