Last updated: October 2, 2026
Tavus Griffin is a new real-time AI model designed for face-to-face human interaction. Tavus says Griffin is its first Human Interaction Model (HIM)—a system that can see, hear, speak, and respond during a live video conversation rather than processing each part of the interaction separately.
In a Tavus-run study, 26 of 54 participants (48%) said they believed they had been speaking with a real person after a one-minute video call with Griffin-Lite. Tavus says this makes Griffin the first model to pass what it calls a real-time video Turing test.
That claim needs context: the study was conducted by Tavus, involved a relatively small sample, and did not include a human control group. So the 48% figure is best understood as a reported result from Tavus’s evaluation—not proof that AI can generally fool people in ordinary video calls.
What Is Tavus Griffin?
Griffin is Tavus’s new approach to conversational AI.
Instead of building an AI video experience as a sequence of separate components—such as speech recognition, an LLM, text-to-speech, and avatar animation—Tavus describes Griffin as a video-to-video model designed to handle perception, conversational decisions, speech, and expressive video generation together.
The model continuously watches and listens while responding.
That means Griffin can potentially:
- Listen while a person is still speaking
- Respond without waiting for a complete conversational turn
- React to interruptions
- Notice visual information
- Change its response based on what it sees
- Produce facial expressions and gestures during conversation
- Maintain a more continuous back-and-forth interaction
Tavus calls this approach a Human Interaction Model because the goal is not simply to generate an AI avatar, but to model the interaction itself.
Griffin Is More Than an AI Avatar
This distinction is important.
Traditional AI avatars generally involve several technologies working together. One system may understand speech, another may generate the response, another may create a voice, and another may animate a face.
Tavus’s Griffin is designed around a different architecture.
The company describes it as a unified video-to-video system in which perception, response timing, speech, movement, and visual generation operate together in real time.
This is part of a broader shift in AI interaction: instead of requiring users to translate everything into typed commands, AI systems are increasingly being designed to understand richer combinations of voice, video, facial expressions, gestures, and environmental context.
For a useful introduction to how AI systems interpret different pieces of information, see Promptenza’s Prompt Anatomy guide.
Why 48% of People Thought Griffin Was Human
This is the statistic that attracted the most attention.
Tavus says it recruited participants in the United States and Europe through an independent research platform.
Participants were told they would have a one-minute video conversation with another participant about what they were looking forward to that year.
They were actually speaking with a PAL powered by Griffin-Lite.
After the conversation, participants were asked whether their conversation partner had been a real person.
According to Tavus:
26 of 54 participants—or 48%—said they believed their partner was human.
For comparison, Tavus says its previous Phoenix-4.5-based system produced a 2.4% result under the same protocol, with 1 of 41 participants believing they had spoken to a real person.
But Does 48% Mean Griffin “Passed” the Turing Test?
Tavus describes Griffin as the first model to pass a real-time video Turing test.
However, readers should understand what that means in this particular experiment.
The study was designed by Tavus and used 54 participants. There was also no comparison group of people speaking to an actual human under the same conditions. Independent reporting has therefore pointed out that the result should not automatically be generalized to all video conversations or treated as a definitive measurement of human-level intelligence.
So there are two separate facts:
Documented result: 26 of 54 participants in Tavus’s study believed Griffin-Lite was human.
Broader interpretation: Whether this constitutes a generally applicable “Turing test” milestone depends on how the test is defined and evaluated.
That distinction matters when interpreting AI benchmark claims.
How Griffin Is Different From Traditional Conversational AI
The biggest change is continuous interaction.
A conventional voice or chatbot interaction often works roughly like this:
You speak → AI listens → AI processes → AI responds
Griffin is designed around a more continuous loop:
You speak + AI watches + AI listens → AI decides whether to respond → AI speaks and reacts
According to Tavus, Griffin can make decisions on very short conversational intervals while continuing to listen and observe. The company says this allows it to respond to interruptions, react to visual information, and avoid some of the unnatural pauses associated with turn-based systems.
Full-Duplex Interaction
Tavus uses the term full-duplex to describe Griffin’s conversational behavior.
In simple terms, full-duplex interaction means the AI does not necessarily have to wait until you completely stop talking before it can react.
For example, during a conversation, an AI could:
- acknowledge something while you continue speaking
- stop speaking when you interrupt
- respond to something appearing on camera
- react to pauses
- maintain conversational timing
This is different from a simple chatbot that waits for a complete message before producing an answer.
Griffin Can See, Hear and Respond at the Same Time
Griffin’s multimodal nature is one of its most important characteristics.
Tavus says Griffin can process the conversation as more than just words.
It can respond to:
- speech
- video
- pauses
- interruptions
- facial behavior
- gestures
- visual context
- conversational timing
This matters because human communication isn’t purely linguistic.
If someone holds an object up to a camera, pauses halfway through a sentence, interrupts the speaker, or changes their expression, a human conversation partner naturally incorporates those signals into the interaction.
Griffin is designed to move AI closer to that type of interaction.
This is where AI Prompting Techniques becomes increasingly relevant. As AI interfaces become more multimodal, interacting with AI may involve more than writing the perfect text prompt. Users may increasingly communicate through voice, images, video, documents, screens, and real-world context.
What Makes Griffin’s Technology Different?
Tavus’s existing technology stack already includes several specialized models.
Its platform describes:
- Sparrow for conversational turn-taking
- Phoenix for real-time human rendering
- Raven for multimodal perception
- Griffin for unified human interaction
Tavus says Sparrow focuses on conversational timing, Phoenix on realistic rendering, and Raven on perception of voice, video, and environmental context. Griffin brings the interaction together into a unified video-to-video model.
From Separate AI Components to Unified Interaction
This reflects an important direction in AI development.
Instead of treating AI as a collection of independent tools, researchers are increasingly exploring systems capable of handling multiple stages of an interaction together.
That could eventually mean AI systems that don’t require users to consciously decide:
“I need to type a prompt now.”
Instead, the interaction could become closer to:
“Show the AI what I’m working on and explain what I need.”
That shift could fundamentally change how people interact with AI.
For more background on how prompting itself is evolving, see Promptenza’s Prompt Engineering guide.
What Can Tavus Griffin Be Used For?
Griffin is still a research preview rather than a broadly available consumer model, so many potential applications remain future-facing.
Tavus describes several possible use cases for this kind of human-computer interaction.
AI Tutors
An AI tutor could potentially watch how a student responds and adjust an explanation based on the conversation.
Instead of simply answering:
“Explain this topic.”
the AI could observe that the learner is confused and change its explanation accordingly.
Customer Support
Real-time video AI could eventually help customers troubleshoot products while seeing what they are showing on camera.
For example, a customer could hold a device up to the camera while explaining the problem.
The AI could potentially combine the visual information with the conversation.
Training and Interview Practice
A conversational AI could simulate interviews, presentations, negotiations, or other real-world conversations.
Because the system can respond to interruptions and conversational cues, the interaction could be more dynamic than a traditional text chatbot.
AI Assistants
The broader vision is an assistant that can interact through natural conversation rather than requiring users to operate a traditional software interface.
Tavus describes this direction as human computing—computers that people interact with more like collaborators than conventional software interfaces.
Griffin Is Not Just About Making AI Look Human
It would be easy to interpret Griffin as simply another photorealistic AI avatar.
That misses the more important development.
The significant change is interaction.
A realistic face is relatively easy to understand.
A system that can:
- see what you’re doing,
- understand what you’re saying,
- decide when to respond,
- interrupt naturally,
- react to visual information,
- generate speech,
- generate facial behavior,
- and continuously adapt during the conversation
is a substantially different type of interface.
That is why Tavus calls Griffin a Human Interaction Model rather than simply an avatar model.
Griffin and the Rise of More Autonomous AI
Griffin also fits into a larger change happening across the AI industry.
AI is moving from systems that simply respond to typed requests toward systems that can understand context, interact with software, use tools, and participate in longer workflows.
Promptenza recently covered this trend in AI Agents Are Becoming More Autonomous.
The connection is important.
AI agents focus on what an AI can do.
Human Interaction Models focus more on how humans interact with AI.
The two trends can eventually converge.
Imagine an AI that can see your screen, understand your voice, communicate through a realistic video interface, browse a website, use tools, and complete tasks while you supervise it.
That would be considerably different from today’s traditional chatbot experience.
What About AI-Generated Video?
Griffin also demonstrates how the boundary between conversational AI and AI-generated video is changing.
Traditional AI video generation generally starts with a prompt such as:
Create a cinematic scene of…
The model then generates a video.
Real-time conversational video works differently.
The AI must generate visual output while simultaneously participating in a conversation.
That means timing becomes critical.
The system isn’t simply creating a finished video. It is continuously producing visual behavior in response to another person.
For readers interested in this broader area, Promptenza’s AI Video Prompting guide covers how prompting for AI-generated video differs from ordinary text prompting.
The Safety Problem: What Happens When AI Looks Too Human?
Griffin’s most interesting feature is also one of its biggest challenges.
If an AI becomes convincing enough that people cannot immediately recognize it as AI, transparency becomes increasingly important.
Tavus itself acknowledges this concern.
The company says the same properties that make Human Interaction Models powerful also create the possibility of misleading people into believing they are interacting with another human.
That is why Griffin-Lite is currently described as a research preview available to select trusted testers rather than a normal customer-facing release. Tavus says it is working on disclosure and safety procedures before a broader release.
Why AI Disclosure Matters
Imagine receiving a video call from someone who appears to be a real person.
You might naturally assume:
- the face is real
- the voice belongs to a person
- the reactions are genuine
- the conversation is happening with another human
As AI-generated video becomes more convincing, those assumptions become less reliable.
This creates an important question for the entire AI industry:
Should AI systems that look and behave like humans always identify themselves as AI?
The technology alone cannot answer that question. Product design, disclosure policies, regulation, and user expectations will all matter.
How Griffin Fits Into the AI Tools Ecosystem
Griffin isn’t an isolated development.
It is part of a rapidly expanding ecosystem that includes:
- AI assistants
- AI agents
- multimodal models
- AI video generators
- AI avatars
- voice AI
- computer-use systems
- real-time AI interfaces
- autonomous workflows
Promptenza’s AI Tools section covers this broader ecosystem and can help readers understand how different categories of AI tools are evolving.
The important shift is that AI is becoming less dependent on a single interface.
The future AI interface could involve text + voice + video + screen + tools + context rather than a text box alone.
What Griffin Means for Prompting
Griffin could also change how we think about prompting.
For years, improving AI results largely meant learning how to write better text instructions.
That remains important.
But multimodal AI is expanding the definition of a “prompt.”
A prompt could increasingly include:
- spoken instructions
- images
- video
- documents
- screen context
- examples
- environmental information
- ongoing conversation
Instead of writing:
“Tell me what’s wrong with this.”
a future AI interaction could involve showing the AI the object, explaining the problem verbally, and allowing it to ask questions in real time.
That is a much richer interaction model.
To understand the components that make instructions clearer and more useful, see Promptenza’s Prompt Anatomy guide.
For definitions of concepts such as multimodal AI, LLMs, AI agents, and related terminology, the AI Glossary is also useful.
Tavus Griffin vs Traditional AI Chatbots
| Feature | Traditional AI Chatbot | Tavus Griffin |
|---|---|---|
| Text conversation | Yes | Potentially |
| Voice interaction | Often available | Yes |
| Live video | Usually limited | Core focus |
| Visual perception | Varies | Designed into interaction |
| Interruptions | Limited/varies | Designed for real-time interruption |
| Facial expression | Usually absent | Generated during interaction |
| Real-time interaction | Varies | Core objective |
| Multimodal interaction | Increasingly common | Central to design |
| Human-like video interaction | Limited | Core focus |
| Availability | Widely available across many products | Griffin-Lite is a research preview |
The table describes the product and technical direction reported by Tavus; availability and capabilities can change as Griffin develops.
Is Tavus Griffin Available Yet?
Not in the same way as a typical consumer AI chatbot.
Tavus says Griffin-Lite is currently a research preview available to select trusted testers, while a broader release is expected after additional safety and disclosure work.
That means the 48% result should not be interpreted as evidence that anyone can currently open a public chatbot and have the same experience.
The technology is being demonstrated and evaluated while Tavus works through the implications of making highly human-like AI interactions more broadly available.
What Happens Next for Human-Like AI?
Griffin points toward an AI interface that looks very different from the traditional chatbot.
Instead of:
Open app → type prompt → receive answer
the interaction could become:
Look at AI → speak naturally → show AI what you’re seeing → AI responds → continue conversation
That could make AI substantially easier for people who don’t know how to formulate technical prompts.
It could also make AI interfaces more accessible because users would not always need to understand the underlying software or prompting syntax.
But the closer AI gets to natural human interaction, the more important transparency becomes.
The technology needs to improve not only at appearing human, but also at making it clear when the user is interacting with AI.
The Bigger AI Trend Behind Griffin
The most important takeaway from Griffin isn’t simply that an AI fooled 48% of participants.
It is that AI interaction is becoming increasingly multimodal, continuous, and context-aware.
We’ve already seen AI move from:
Text → Voice → Images → Video → Agents → Real-time interaction
Griffin represents another step toward systems that can combine several of those capabilities in one conversation.
That makes this development relevant far beyond AI avatars.
It could eventually affect:
- education
- customer service
- healthcare interfaces
- professional training
- accessibility
- entertainment
- productivity
- AI assistants
- autonomous AI agents
The next generation of AI may therefore be defined not only by how intelligent the underlying model is, but by how naturally people can interact with it.
Frequently Asked Questions About Tavus Griffin
What is Tavus Griffin?
Tavus Griffin is a real-time Human Interaction Model designed to see, hear, speak, and respond during face-to-face video conversations. Tavus describes it as a unified video-to-video system rather than a collection of separate conversational components.
Did 48% of people really think Griffin was human?
According to Tavus, yes. In its study, 26 of 54 participants (48%) believed they had been speaking with a real person after a one-minute conversation with Griffin-Lite.
Did Griffin pass the Turing test?
Tavus says Griffin was the first model to pass its real-time video Turing test. However, this was a company-run study with 54 participants and no human control group, so the result should be understood within the limits of that particular evaluation.
Is Tavus Griffin publicly available?
Griffin-Lite is currently described by Tavus as a research preview available to select trusted testers rather than a general customer release.
What makes Griffin different from an AI avatar?
Traditional AI avatars can combine language models, speech systems and avatar-generation technology. Griffin is designed as a unified real-time video-to-video interaction model that handles perception, conversation and expressive output together.
Can Griffin see what a person is doing?
Tavus says Griffin can react to visual information during a conversation, allowing the model to incorporate what it sees alongside spoken interaction.
Why is Griffin important for AI?
Griffin demonstrates a move toward AI systems that communicate through continuous, multimodal interaction rather than relying primarily on typed prompts. It could influence how people interact with AI assistants, agents and other AI-powered applications.
Test Your AI Knowledge
1. What is Tavus Griffin?
A. A text-only LLM
B. A Human Interaction Model
C. An image-generation model
D. A search engine
Answer: B
Griffin is described by Tavus as its first Human Interaction Model.
2. What percentage of participants in Tavus’s study believed Griffin-Lite was human?
A. 12%
B. 24%
C. 48%
D. 75%
Answer: C
Tavus reported that 26 of 54 participants, or 48%, believed their conversation partner was a real person.
3. What does “full-duplex” mean in Griffin’s context?
A. The AI only communicates through text
B. The AI can listen and respond continuously rather than waiting for rigid turns
C. The AI generates two videos simultaneously
D. The AI translates two languages
Answer: B
Full-duplex interaction allows the system to continue listening and responding during a live conversation.
4. What makes Griffin different from a conventional chatbot?
A. It only works offline
B. It focuses on continuous face-to-face multimodal interaction
C. It cannot process voice
D. It only generates images
Answer: B
Griffin is designed around real-time video interaction rather than text-only conversation.
Final Takeaway
Tavus Griffin is significant because it shifts the AI conversation from generating answers to participating in interaction.
The headline statistic is striking: 48% of participants in Tavus’s study believed Griffin-Lite was a real person after a one-minute video call.
But the more important development is underneath that number.
Griffin is designed to see, hear, speak, react, interrupt, and respond within a continuous face-to-face interaction. That represents a different approach to AI interfaces—one where the user may not need to formulate every request as a carefully written prompt.
At the same time, the technology raises an important issue: the more convincing AI becomes, the more important clear disclosure becomes.
For Promptenza readers, Griffin is another sign that the future of prompting may extend well beyond typing instructions into a chatbot. AI interaction is becoming increasingly multimodal, contextual, visual, and conversational.
And that means learning how to communicate with AI will remain important—even as the interface itself changes.
Last updated: October 2, 2026
