Archive · 3 Oct 2026
Tavus Says Griffin Passes Video Turing Test With 48% Success
Tavus has introduced Griffin, an AI designed for live video conversations.
The company says 48% of people who spoke with it thought it was a real person.
Griffin can listen and watch while it talks, and it can react to pauses or interruptions.
It also generates speech, movements, and changes to the scene in real time.
Tavus says Griffin can make gestures and change backgrounds using one reference image.
A smaller version called Griffin-Lite is available to some developers.
Tavus says it plans to release the full model more widely in the coming months.
Some critics have raised concerns about possible misuse, while Tavus says it is working on safeguards and disclosure.
Tavus says 48% of people who spoke with Griffin live thought it was a real person.
The company describes Griffin as a real-time Human Interaction Model that can perceive and respond to audio and video.
Its two-engine architecture handles conversational timing and generates speech, movement, and scene adjustments.
Tavus says Griffin can create dynamic scenes from one reference image, including gestures, shadows, and background changes.
A limited developer preview, Griffin-Lite, is available; Tavus says a broader release is planned for the coming months.
- Who
- Tavus and its Griffin model; Tavus founder Hassaan Raza also commented on the announcement.
- What
- Tavus says Griffin passed a video Turing test, with 48% of live participants mistaking it for a human.
- When
- Tavus announced Griffin on October 1, 2026; Hassaan Raza posted a follow-up on Saturday.
- Where
- Tavus announced the model on X; the test location is not stated.
- Why
- Griffin is designed to hold real-time face-to-face conversations by interpreting and responding to audio and video as they unfold.
Concerns about AI impersonation and misuse
Tavus's case for development with safeguards
Whether human-like AI should advance
Concerns about AI impersonation and misuse
Critics, including U.S. Senator Bernie Sanders, condemned the prospect of AI appearing human and called for a pause; some commenters warned it could be used to scam older people.
Tavus's case for development with safeguards
Tavus says Griffin is a limited research preview and argues that sharing the work openly can accompany development of a safe public release.
Disclosure and safety
Concerns about AI impersonation and misuse
Public reaction described in the article reflects suspicion about how a convincing AI video presence could be misused.
Tavus's case for development with safeguards
Tavus says it is working with partners on safeguards and disclosure and inviting discussions with officials about regulation and safe use; founder Hassaan Raza said people will know when they are interacting with AI.
- Reported result
- Tavus says 48% of people who spoke with Griffin live thought it was a real person.
- Model description
- Tavus calls Griffin its first Human Interaction Model (HIM).
- Architecture
- A Continuous Conversational Modeling engine and a unified Audio-Visual Generation engine.
- Capabilities described
- Real-time speech, movement, scene adjustments, and responses to pauses or interruptions.
- Scene generation
- Tavus says Griffin can generate dynamic scenes from a single reference image.
- Preview
- Griffin-Lite is available to a select group of developers.
- Planned release
- Tavus says broader availability of the full model is planned for the coming months.
- Concerns and safeguards
- Critics raised concerns about misuse; Tavus says it is working with partners on safeguards and disclosure.
