Tavus Griffin AI Model Claims to Pass Video Turing Test With 48% of Users Mistaking It for a Human

Tavus has unveiled Griffin, a real-time AI video model designed for human-like conversations, claiming 48% of participants mistook its responses for those of a real person.

Tavus Introduces Griffin for Human-Like Video Conversations

AI company Tavus has introduced Griffin, a new model designed to make real-time video conversations with artificial intelligence feel more natural and human-like.

The company describes Griffin as an end-to-end Human Interaction Model (HIM) capable of handling face-to-face video interactions in real time.

Unlike conventional conversational AI systems that combine separate technologies for speech recognition, language generation and video creation, Griffin is designed to process and generate video interactions through a unified pipeline.

Griffin Claims 48% Video Turing Test Success

According to Tavus, 48% of people who interacted with Griffin in live conversations believed they were speaking with a real human.

The company presented the result as evidence that its system can reproduce several characteristics associated with natural human communication.

The claim comes in the context of the Turing test, which broadly examines whether a machine can produce behaviour that people cannot reliably distinguish from that of a human.

Tavus says Griffin represents a significant step toward making AI interactions more fluid and less recognisably artificial.

Single Video-to-Video AI Pipeline

Traditional conversational AI systems often operate through multiple stages. A user's speech may first be converted into text, processed by a language model and then transformed into synthetic speech and video.

Griffin is designed differently. Tavus says the model operates as a full-duplex video-to-video system, allowing it to simultaneously perceive and respond to visual and audio information.

This means Griffin can see, hear, interpret and respond during an ongoing conversation rather than waiting for each individual stage to finish.

AI Can React to Interruptions and Visual Changes

One of Griffin's key features is continuous perception while the model is speaking. According to Tavus, the system can respond when a user interrupts, pauses, changes their expression or introduces a new visual cue.

This allows the model to adjust its responses while a conversation is taking place instead of following a rigid question-and-answer sequence.

The company says this approach is intended to make conversations feel more dynamic and closer to natural face-to-face interaction.

Griffin Uses Two Core Engines

Tavus says Griffin's architecture includes two major components designed to manage different aspects of real-time interaction.

The Continuous Conversational Modeling engine continuously evaluates incoming information at sub-second intervals. It helps determine when Griffin should respond, pause or use conversational back-channel signals such as brief acknowledgements.

The second component is the Audio-Visual Generation engine, which produces speech, physical movements and other visual elements during the interaction.

Together, the systems are designed to coordinate conversation and visual behaviour in real time.

Real-Time Facial and Body Movements

Griffin is designed to generate more than basic facial animation. Tavus says the model can create complete dynamic scenes from a single reference image, including movements involving the hands and arms.

The system can also adjust elements such as shadows and backgrounds while a conversation is taking place.

This approach is intended to make AI avatars appear more responsive and physically expressive rather than simply displaying pre-generated facial movements.

Griffin Understands Conversational Timing

Another feature highlighted by Tavus is Griffin's ability to account for the timing of conversations. The model can analyse how long a person remains silent and use that information to determine what may be happening in the interaction.

For example, a pause could indicate that someone is thinking about their response, or it could mean they are waiting for the AI to continue.

Understanding such pauses is important for creating more natural conversations because human dialogue is not limited to spoken words.

Griffin-Lite Available to Selected Developers

Tavus has released an early research preview called Griffin-Lite to a limited group of developers. The company plans to expand access to the technology and introduce the full Griffin model more broadly in the coming months.

The limited release will allow developers to explore the model's capabilities and provide feedback before a wider rollout.

AI Video Conversations Move Beyond Traditional Chatbots

The development reflects a broader shift in artificial intelligence from text-based chatbots toward systems capable of interacting through multiple forms of communication.

Real-time video models could potentially be used in areas such as virtual assistants, customer service, education, digital humans and interactive applications.

However, Tavus' reported 48% result represents the company's own announcement, and the methodology and testing conditions behind the figure are important when assessing what the result demonstrates.

What Griffin Means for AI Interaction

Tavus' Griffin aims to make AI conversations more responsive by combining perception, language, audio and visual generation into a real-time system.

Its reported performance in live interactions highlights the growing ability of AI models to reproduce aspects of human conversation and behaviour.

As Griffin-Lite reaches more developers and the full model becomes available, broader testing will provide more information about how consistently the technology can perform in real-world video conversations.