Griffin by Tavus: The First AI to Pass the Video Turing Test (48% Fooled)

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Griffin by Tavus: The First AI to Pass the Video Turing Test (48% Fooled) · Channel: Prompt Engineer 48 · 16:31 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI safety

Summary

The video introduces Griffin, Tavus’s real-time video-to-video model for face-to-face interaction, and presents demo clips showing its emotional responses, full-scene generation, interruption handling, visual perception, and sensitivity to timing. It explains how Griffin replaces a turn-based chain of speech and video systems with a model that continuously processes audio and video and generates speech and imagery as the conversation unfolds. The presenter reviews latency and benchmark results, including a study in which 48% of participants identified Griffin as a real person after a one-minute call. The video also highlights the risks of a system that can be mistaken for a human: Griffin Light is limited to trusted testers while Tavus works on disclosure and safety measures.

Key points

Griffin is presented as a human interaction model designed to understand and generate face-to-face interaction in real time.
The demos show Griffin reacting to emotion and context, generating the full scene, handling interruptions, interpreting visual information, and responding to silence and timing.
Unlike a turn-based cascade, Griffin processes incoming audio and video continuously and decides at sub-second intervals whether to speak, react, or wait.
Its speech system generates audio progressively, while its video system produces 720p video in streaming chunks.
The video reports an average latency of 0.43 seconds on H100s for Griffin Light.
On Nvidia’s Video of DB benchmark, Griffin scored 3.83 out of 5 on generation and 3.73 on perception, according to the video.
In a study of 54 one-minute calls, 26 participants (48%) said Griffin was a real person.
The video notes that the same qualities that make Griffin feel natural could mislead people about whether they are speaking with AI.
Griffin Light is described as a research preview for selected trusted testers, with safe-disclosure features still in development.
The video also points to existing Tavus tools for building face-to-face AI experiences while Griffin is not yet on the TEOs platform.

Chapters

Start	Chapter	Summary
00:00	Griffin demo and introduction	The video opens with a live conversation with Griffin and introduces Tavus’s claims about the model.
03:06	Why Griffin was built	The presenter explains the limitations of turn-based AI and Griffin’s continuous, full-duplex approach.
04:41	Interaction capabilities in the demos	A series of clips demonstrates emotional responses, full-scene generation, interruptions, visual perception, and temporal understanding.
10:51	Architecture and generation	The video outlines Griffin’s conversational and audio-visual generation engines, along with its streaming speech and video methods.
12:54	Benchmarks and face-to-face study	The presenter reviews reported latency, benchmark scores, and results from one-minute calls with participants.
14:58	Safety, access, and future uses	The video discusses disclosure concerns, Griffin Light’s limited availability, potential applications, and Tavus’s current tools.
16:31	Closing	The presenter recaps the video’s topics and directs viewers to the links in the description.

Quotes

48% of people believed Griffin was a real person after a 1-minute video call. — 02:03

What Tava’s wants is for computing to become invisible — 03:38

It never stops thinking. — 04:10

26 of them, that’s 48% said it was a real person. — 14:27

The same properties that make this a great interface also let it deceive a person into believing it is not AI. — 14:58

Concepts and entities

Griffin, Griffin Light, Tavus, Hassan Raza, Ioannis Patras, Tavic, VDT, Phoenix, Raven, Sparrow, NVIDIA, teste de Turing por vídeo

Full transcript (original language)

00:00 All right. In this video we’re going to look at Griffin, the new model from Tavus. And you can see here that this is a live video call and the woman on the laptop screen is not a person. Every pixel of her is generated by Griffin in real time. Let’s go ahead and watch the full conversation first. >> All righty, let’s do this. >> Hey Hassan, how’s it going? It’s pretty late for you. Have you thought any more about how we’re introducing Griffin? >> Um I don’t know. What are your ideas?

00:30 >> I mean I think we should stop trying to explain it and just, you know, show them what it looks like. >> Do you think that we should just show a recording of us talking or something? >> I think a video of just us like this would be the best way to actually prove it. >> Do you think they’ll believe that you’re AI? >> I think that’s the whole point, isn’t it? If it’s good enough, they won’t even have to think about it. >> Now, how about just a thumbs up? >> Sure, one thumbs up coming right up. >> Hell yeah. That was awesome. I would say

01:02 you’re my favorite coworker. [music] It’s been It’s been really great to uh work on this with you. >> You’re making me blush. Aw, who’s that with the ugly sweater? >> Yeah, I gifted this. >> Yeah, it it definitely looks like it. >> Oh [laughter] my gosh, okay. >> 1%? You’re living on [music] the edge, Hassan. Truly. I’ll keep my breathing steady so I don’t drain your CPU any faster. >> Wow, you actually did slow down your breathing. >> It’s like the line between talking to a machine and just

01:33 talking to someone is getting really, really thin. >> Yeah, it really feels like we’re like, you know, passing a Turing test soon. >> But honestly, if we want people to see the human part of the simulation, maybe that’s the most real thing we could show them. What do you think? Is it too much? >> No, no. I um >> So that was Hassan, the Tavus CEO, talking to Griffin. Nobody is reading from a script there and near the end he says the line between talking to a machine and just talking to someone is

02:03 getting really, really thin. Now, that is a big claim, so let’s go through the page properly. Griffin is the world’s first human interaction model or HIM. It’s a new class of model designed to understand and generate face-to-face, real-time human interaction. It listens while it talks and it pays attention to expressions and pauses, not just words. Here are the three numbers from the top of the page. 48% of people believed Griffin was a real person after a 1-minute video call. It’s number one on Nvidia’s independent test of face-to-face AI and it’s 37% ahead of

02:36 the next best AI at reacting in the moment. It was announced on October 1st, 2026 in San Francisco by Hassan Raza, the co-founder and CEO, and Ioannis Patras, the head of research. And mind you, what we have today is Griffin light, a research preview with a more powerful model to follow. So, here’s the plan. First, why they built this. Then we’ll play the demo clips from the page one by one. After that, how it’s built, the benchmarks, the safety side, and what’s next. Let’s start with the why. In his message, Hassan says, “Humans are

03:06 designed to communicate face-to-face. We speak as much through expressions, tone, gestures, and timing as we do through words. Machines don’t understand that. They make us learn their language, the right command, the right button click, the right prompt.” And then he lists the little moments that make talking to AI tiring. You share something personal and the face keeps smiling or a cold machine says, “I’m sorry to hear that.” You pause to think and it jumps in on incomplete context or it needs time to answer and nothing in its face tells you it’s working on it. Each of those moments, he says, is you managing the

03:38 machine. You adjust your timing, you simplify your language, you remember not to leave dead air. What Tava’s wants is for computing to become invisible so that in a great conversation, you aren’t thinking about the conversation at all. Now, why does this happen? Most real-time AI works as a relay and they call it a cascade. One system transcribes your speech, a language model replies to the transcript, and other systems turn that reply back into a voice and a face. Every handoff adds delay and throws away information like your tone of voice or what’s on camera.

04:10 Griffin replaces that relay with one full-duplex video-to-video model, perception deciding when and how to respond, and expressive speech and video generation all happen at the same time. It never stops thinking. Every sub-second mini turn, it decides whether to speak up, react, or wait. The page shows four things this makes possible. It says, “Um, hm.” While you’re still talking, it answers without leaving dead air. It stops the moment you cut in, and it reacts to something it sees. Let’s see all of that in the clips. First

04:41 capability, behavior and emotion modeling. Griffin has learned to express emotions and gestures according to the context. It laughs, changes its tone, and shifts in response to what the other person says. In this clip, Rudy tells Vanessa he’s been promoted. Listen to how she reacts before he’s even finished. >> Hm, the promotion. Don’t leave me hanging. >> So, okay. >> get it? >> Basically, yes. I got it. >> Oh my god, yes. That’s amazing. >> Yeah, I’m I’m so glad. >> All those months of grinding finally

05:11 paid off. >> I know. And now one of the best things, I don’t have to work with Vikram anymore. He’s he’s out of my life. I can finally move Second capability, full scene generation, and this one is really amazing. Griffin generates every pixel in every frame in real-time from one reference image, not just the face, arms, and fingers, but the movement of the chair, the shadows, and the background behind them. Let’s go ahead and play a round of Simon Says. >> Okay. >> Wait, what? >> [laughter] >> Okay, Simon says wave your hand. >> I’m doing it.

05:41 >> Oh, nice. Simon says touch your face. >> Okay, here I go. Can you touch your chin? Wait, you didn’t say Simon says. I’m not going to fall for that one, Vanessa. No, nice try. >> So, you can see that she waves, she touches her face, and when the person doesn’t say Simon says, she calls their bluff. Every movement is generated on the spot, body and background included. Third capability, full duplex conversation. Griffin keeps evaluating the state of the conversation

06:12 independent of turns. It can interrupt, backchannel, or be interrupted without losing its place. Here, Ari and Griffin make up a story together, and they cut each other off as they go. >> Hey, Vanessa. So, I’m working on this short story, right? And I’m thinking about having it be like a political drama between two different colonies on Mars. >> Drama? That’s a lot of heavy lifting for a short story. Are we talking like high-stakes government espionage, or just people arguing in a boardroom? >> No, no, high-stakes espionage, Mars

06:43 colonization. We’re fighting over resources, but I don’t know like why why the two factions are really fighting. >> is the missing piece? >> You got it. >> fighting over So, if it’s just generic resource >> Notice that Griffin pushes back, it asks if this is government espionage or just people arguing in a boardroom, and then it suggests one faction treats the resource as sacred while the other other side just burns it. Nobody waited for a turn to end. Fourth capability, perception. Griffin goes beyond words and reacts to visual context when it

07:14 makes sense. In this clip, it coaches Ari through a Rubik’s Cube. It watches the cube as he turns it, keeps acknowledging him while he works, and when he pauses mid-solve, it waits until he’s ready. >> Hey, Vanessa. Um I got this Rubik’s Cube that I’m trying to solve. I’m using the classic like three-layer strategy. I got the first two layers, but this top blue layer, I only have a bar. >> Okay, I’ll see what we’ve got. I’m going to start with the cross sequence. To keep things easy, let’s do this in little batches.

07:44 First batch, one. Turn the side facing you 1/4 turn clockwise. Two. Turn the side on your right 1/4 turn clockwise. Three. Turn the side facing up 1/4 turn clockwise. Let me know when you’ve done those. >> Yep. I’m ready for the next. Oh, nice. I got a cross. Cool. >> Nice. Let me see. >> So, you can see the pacing. It gives the

08:15 moves in little batches and waits for him to say he’s ready. Then it sees the blue cross and says it’s looking good. Let’s skip to the end and see how it finishes. >> Side facing up 1/4 turn clockwise. >> Oh! Yo! >> Now, show me the top face so I can see how the edges are looking. >> Vanessa, we did it. >> Oh my gosh, you actually did it. >> Here’s another one. A teacher is practicing a storytelling game where students show props to put in the story. In the middle of it, someone holds up a toy and Griffin writes it into the plot

08:47 in its very next line without being asked. >> Whenever you’re ready, just kick it off. >> Okay, great. >> I think that we’re in the ocean and there’s a school of fish on a secret mission and they have a map, but maybe they don’t know. >> mission? They’re not just swimming, they’re actually they’re the elite underwater intelligence agency and they’re trying to >> And maybe they actually look like this. >> One more and this one is subtle. Nobody

09:17 mentions it, but Britney looks cozy and Griffin says so. So, it noticed the setup and brought it up the way a friend would. >> Hey Ben, how’s it going? >> It’s It’s going pretty good, honestly. Just, you know, catching up on some things around the house. How about you? You look like you’re in a cozy little setup there. >> Yeah, now I’m super cozy here. I’m just thinking about what I should make for dinner. >> Last capability, temporal understanding. Griffin understands time as part of the conversation. How long a silence has lasted, what a sustained silence means,

09:50 and when to speak again on its own. Sagar is soldering a motherboard and the iron needs about 12 seconds to heat up. >> Okay, I’m done screwing it back together. What’s next? >> Okay, so the next step is to solder the loose connector back onto the motherboard. >> [sighs] >> Okay, let me go let me go grab that, but beforehand, yeah, I turned the soldering iron on. I think it’s going to take about 12 seconds, so maybe just let me know when that’s done. >> Sounds good. I’ll keep an eye on it and let you know

10:20 the second it hits that 12-second mark. Hey, it’s been 12 seconds, it should be ready to go. >> Cool, on my way. Um, I >> And you can see it did what it said. It kept an eye on the clock and spoke up right at the 12-second mark without being asked. That kind of timing is really hard for a turn-based system. Okay, so how is it built? Griffin is a two-part system. First, a continuous conversational modeling engine, which

10:51 perceives the incoming audio and video, decides when and how to respond, and produces signals for what should be said and how. Second, an audio-visual generation engine with streaming speech generation and streaming video generation. The conversational model makes decisions at regular sub-second intervals, not once per turn. At each one, [clears throat] it looks at what’s been said and at the person’s verbal and nonverbal behavior, and decides what to say next and how. The output controls emotional tone, stance, facial expression, and gesture. So, a pause for

11:23 thought is not treated as the end of the turn. The page has a nice diagram comparing the two. The turn-based cascade waits for you to finish, then runs ASR, then the LLM, then TTS, then the avatar, and about 1 and 1/2 seconds of silence go by before the face moves. With Griffin, the reply starts before you finish, and mind you, the page says these timings are illustrative, not measured from a real session. For speech, Griffin uses a fast auto-regressive diffusion transformer, which they call VDT. It conditions on a

11:53 10-second speaker clip and generates speech progressively as control signals arrive. A causal decoder emits audio packets as small as 10 milliseconds. So, it starts talking before the sentence is finished, and it can clone a voice from about 10 seconds of audio. Underneath that is a codec called Tavic. It maps 48 kilohertz audio into a continuous latent, 40 values per frame, 100 frames per second, and no codebooks. That’s 6,000 latent frames per minute against 2.88 million raw samples. Small and continuous, so the model can predict it

12:24 faster than real time. Then the video side, earlier Tavus models relied on 3D priors, which meant no big gestures and no dynamic backgrounds. Griffin generates 720p video in 320 millisecond chunks in real time. One latent is eight frames at 25 frames per second with three diffusion steps per latent from a single reference photo plus streaming audio and controls. To get there, they distilled a big slow many-step diffusion model in three stages. One, distribution matching distillation into a few-step student.

12:54 Two, teacher forcing to make it auto-regressive so it can stream. Three, self-forcing so long conversations don’t drift. That’s how you get something that is both good and fast. Now the evaluations. They tested at three levels, the video generator on its own, the full system on Nvidia’s video of DB benchmark, and live 1-minute video calls with people who didn’t know they were talking to a model. First, latency. That’s the time from audio arriving to the video showing its effect on H100s. Griffin light averages 0.43 seconds,

13:26 which is half that of the next fastest method. Because it generates a single latent with no low-kick audio on quality, it’s first on Dover, first on FID, and first on THEVAL, and second on lip-sync LSEC at 7.27. And I like that they’re honest about that last one. They say LSEC rewards exaggerated mouth movement, which THEVAL accounts for. Then Nvidia’s Video of DB, which Nvidia ran independently. On the generation track, Griffin scored 3.83 out of five.

13:57 The next best system, Gemini 2.5 plus Anam, scored 2.80, and the human reference is 3.92. So, that’s just 0.09 below a human. On the perception track, Griffin scored 3.73. Mini CPM Plexo 4.5 got 3.44. Gemini 2.5 Flash Native got 3.17, and OpenAI’s GPT Real Time got 2.97. The human reference is 4.20. That’s the most perceptive of 15 models. And on takeover rate alignment, it’s the best on both tracks. And now the number

14:27 from the title in the face-to-face study, 54 people had a 1-minute video call with a partner they were told was another participant. 26 of them, that’s 48% said it was a real person. On the previous TAVAS system, Phoenix 4.5 with Sparrow 2 and Raven 1, it was 1 out of 41. That’s 2.4%. Some details from that study, over half said it never crossed their mind that it might be AI. The ones who did suspect usually suspected within the first 20 seconds. And on a 7-point scale, people rated it 5.4 for natural,

14:58 5.6 for trustworthy, and 5.8 for wanting to talk to it again. The lowest was flowing naturally at 4.9, and I like that they report that, too. Now, the part I think matters most. The page says that since Griffin Light is the first model to pass the Turing test, they understand the responsibility to be thoughtful. The same properties that make this a great interface also let it deceive a person into believing it is not AI. So, Griffin light is not available to customers right now. It’s available to select trusted testers as a

15:28 research preview, and you can submit a form to request access. They’re working on safe disclosure features and with organizations tackling AI safety, and they expect to release Griffin very soon after these concerns are addressed. What’s next? They give three examples. A student gets a tutor who notices when an explanation isn’t making sense and tries another one. Someone at work rehearses a hard conversation with a counterpart who reacts the way a real person might, and a customer holds a broken part up to the camera and works out the fix together

16:00 without knowing what the part is called. Now, if you want to build today, Griffin isn’t on the TEOs platform yet, but you can build PALS AI you talk to face-to-face on the models that 150,000 developers and businesses already use. Phoenix generates the face, Raven sees and understands you, and Sparrow knows when to talk. There’s the TEOs API, and there’s PAL Maker, a no-code way to build a PAL. Credits go to the TEOs research and engineering teams with thanks to Bastion, Daily, and Cerebrium

16:31 for the infrastructure, Queen Mary University of London for research collaboration, and Nvidia for building and scoring the benchmark. So, in this video we have seen Griffin from TEOs, the first human interaction model. We went through the clips, the architecture, the benchmarks, and the safety side. The links are in the description, so go ahead and check them out. And let me know what you think. I’ll see you in another one.