Tavus Griffin: Did an AI Video Avatar Really Pass the Turing Test?
Tavus says its Griffin model fooled 48% of people on video calls. Here's what the claim means, what full duplex AI is, and how solid the test is.
What is Tavus Griffin?
Tavus Griffin is an AI video avatar model built by Tavus, a company that specializes in real-time video AI. Tavus says Griffin is the first model to pass a “video Turing test,” meaning that in live video conversations, a meaningful share of people couldn’t tell whether they were talking to a human or a machine. In Tavus’s own study, 48% of participants believed Griffin was a real person. The company says its previous systems scored under 3% on the same kind of test.
TL;DR
- Tavus Griffin is a real-time AI video avatar that Tavus claims passed a video version of the Turing test, with 48% of people in a study mistaking it for a human.
- The test involved 54 people having one-minute conversations with Griffin, which makes it a small, company-run study rather than independent third-party research.
- Griffin runs on full duplex processing, meaning it listens, watches, thinks, and generates speech and facial movement simultaneously instead of waiting for a pause to respond.
- Tavus ranks Griffin number one on Nvidia’s benchmark for full duplex AI video, giving the claim at least one outside reference point beyond Tavus’s internal numbers.
- Tavus is branding this as a new category called a “human interaction model,” distinct from standard conversational AI or video generation tools.
- The timing matters because voice AI is also getting dramatically faster, with Microsoft shipping new real-time voice and transcription models the same week, suggesting the entire stack for lifelike AI conversation is maturing at once.
- Whether Griffin’s result holds up under scrutiny depends heavily on sample size, conversation length, and who’s doing the testing, all of which favor Tavus in this particular study.
One coffee. One working app.
You bring the idea. Remy manages the project.
How does Tavus Griffin actually work?
Griffin is built for full duplex interaction, which is the technical detail that separates it from most chatbots and voice assistants. A typical voice AI works in turns: you speak, it waits for silence, then it processes your words and replies. That turn-taking creates the slightly stilted rhythm most people associate with talking to a machine.
Full duplex systems do away with that structure. Griffin is simultaneously seeing the person on camera, hearing their speech, deciding in real time when and how to respond, and generating its own synthesized speech and facial animation, all at once. That lets it interrupt, react mid-sentence, and hold the kind of overlapping, messy back-and-forth that real human conversation actually has. That overlap and responsiveness is likely a big part of why some participants in Tavus’s study didn’t immediately clock it as artificial.
Tavus is calling this a new category of AI: a “human interaction model,” distinguishing it from large language models focused on text generation or standard video avatar tools that just animate a face over pre-scripted audio.
What exactly is the “video Turing test,” and did Griffin really pass it?
The original Turing test, proposed by Alan Turing, asks whether a human evaluator can distinguish a machine from a person based on conversation alone. A video Turing test applies the same logic to real-time video calls: if someone on a video call can’t tell they’re talking to an AI, the AI passes.
Tavus ran this test with 54 people, each having a one-minute conversation with Griffin. Afterward, participants were asked whether they believed they’d spoken to a human or an AI. 48% said human. Tavus contrasts that with its own older systems, which it says scored under 3% on the same measure.
It’s worth being precise about what this does and doesn’t prove. A 48% fool rate across 54 people in one-minute conversations is a meaningful signal, but it’s a small sample run and evaluated by the company that built the product. There’s no independent lab, academic peer review, or third-party replication mentioned alongside it. One-minute conversations also limit how deeply a person can probe for inconsistencies, slow reactions, or visual glitches that might surface in a longer exchange. None of that means the result is wrong, but it does mean “passed the Turing test” should be read as “performed strongly in Tavus’s own short-form test,” not as an independently verified, peer-reviewed finding.
The one external data point in the mix is that Griffin is ranked number one on Nvidia’s benchmark for full duplex AI video, which at least gives the claim a reference point outside of Tavus’s internal messaging.
Why does full duplex video AI matter right now?
Full duplex matters because it’s the missing piece that’s kept AI avatars and voice agents feeling obviously synthetic. Text chatbots solved fluent language. Voice models have been racing to kill latency and turn-taking awkwardness. Video avatars have mostly been the weakest link, often looking like a static face with lip movement bolted onto an audio track.
Griffin folding vision, hearing, decision-making, and generation into one simultaneous process is a sign that the avatar side of the stack is catching up to where voice and text already were. That has obvious implications for things like AI customer support reps, virtual assistants on video calls, sales and training simulations, and any product where a human-like presence on screen adds value over a plain chat window.
The timing lines up with a broader push across the industry toward faster, more natural real-time interaction. The same week Tavus put out Griffin, Microsoft released new voice models, including a streaming transcription model and two text-to-speech models built for low latency, aimed at cutting out the awkward pauses that make automated phone systems and voice agents feel robotic. Different companies, different modalities, same underlying goal: make AI conversation feel less like querying a machine and more like talking to someone.
Is the “human interaction model” category a real shift or just marketing?
Tavus describing Griffin as a new category, a “human interaction model,” is partly a technical claim and partly positioning. Technically, full duplex video generation that handles perception and generation concurrently is a genuine capability jump from earlier avatar systems that were effectively doing turn-based video dubbing. That’s a real architectural difference worth naming.
Whether it deserves to be treated as its own category of AI model, separate from conversational AI or video generation more broadly, is less settled. Categories like this tend to get defined by adoption and competition, not by the company that coins the term. If other labs ship comparable full duplex video avatars and the term sticks across the industry, it’ll have earned the label. If Tavus remains the main company using the phrase, it’s more branding than taxonomy. Right now, with one company’s internal benchmark behind it, it’s too early to say which way that goes.
Frequently Asked Questions
What does “full duplex” mean in AI video?
It means the AI processes input (seeing and hearing you) and generates output (speaking and animating a face) at the same time, rather than waiting for you to finish talking before responding. This allows for interruptions and more natural, overlapping conversation.
How many people actually tested Tavus Griffin?
Tavus’s study involved 54 people, each having a one-minute real-time video conversation with Griffin, after which they were asked whether they thought they’d spoken to a human or an AI.
Is Tavus Griffin’s Turing test result independently verified?
Not based on available information. The 48% figure comes from a study Tavus ran and reported itself. The only outside reference point mentioned is Griffin’s number one ranking on Nvidia’s benchmark for full duplex AI video, not an independent replication of the Turing test itself.
How does Griffin compare to Tavus’s earlier models?
Tavus says its previous systems scored under 3% on the same fooling-rate measure where Griffin scored 48%, suggesting a large jump in how convincingly its avatars can pass as human in short conversations.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Why are companies racing to build more human-like AI video and voice right now?
Faster, more natural real-time AI interaction (across voice, video, and text) broadens how AI can be used in areas like customer support, virtual assistants, and training simulations, where a lifelike presence matters more than a plain text interface.