Tavus Griffin passes video Turing test

The 48% headline comes from a specific, bounded test: a blind study where 54 participants had one-minute video calls, 26 of them (48.1%) concluded they'd spoken to a real person, versus just 1 of 41 participants (2.4%) who were fooled by Tavus's prior generation of models built from separate rendering, perception, and conversation components ^[1]. On NVIDIA's independent VideoFDB benchmark, which evaluates full-duplex AI video systems, Griffin-Lite ranked #1 of 15 models on both the generation and perception tracks, scoring 3.83 out of 5 on generation against a human baseline of 3.92, more than a full point ahead of the next-best system, a combination of Gemini 2.5 and Anam ^[1]. Tavus also reported audio-to-video latency of 0.43 seconds on H100 chips, about 50% faster than the next-closest system, which matters because any lag between what a person says and how the avatar reacts is one of the easiest tells that something isn't human ^[1].

The reason Griffin can close that gap isn't a bigger version of the same approach, it's a different architecture. Most conversational video avatars chain together separate systems: one that perceives audio and video, one that decides what to say, and one that renders the face and voice. Griffin is a single full-duplex, video-to-video model that perceives, decides, and generates at the same time, instead of passing data through that pipeline in sequence ^[2]. It generates 720p video in 320-millisecond chunks and streams speech in packets as small as 10 milliseconds, and it can clone a voice from about 10 seconds of reference audio ^[2]. Tavus built Griffin by consolidating three models it had already shipped separately - Phoenix-4.5 for rendering, Raven-1 for multimodal perception, and Sparrow-2 for conversational dynamics - into one architecture ^[3]. That consolidation follows a fast funding run: the company raised a $40 million Series B led by CRV in November 2025, bringing its total funding to roughly $64-70 million since a 2023 seed round, on top of a platform the company says is already used by more than 150,000 developers and businesses ^[5][6]. The clearest illustration of what the single-model approach buys is a demo where a person solders a motherboard while Griffin tracks his progress visually and speaks up when the next step is due, rather than only responding once he stops talking ^[4].

Tavus has been unusually direct about the downside of its own achievement. In its own materials, the company wrote that the same properties that make Human Interaction Models powerful interfaces for natural communication between human and machine also allow them to deceive a human into believing it is not AI ^[2]. That's not hypothetical: the company pointed to a 2024 case in Hong Kong where staff at the engineering firm Arup were tricked into transferring about $25 million after a video call with deepfaked colleagues, as the kind of risk a model this convincing has to reckon with ^[2]. There's also a regulatory angle: the EU AI Act's Article 50 requires that people be told when they're interacting with an AI system unless it's obvious, a rule that becomes much harder to satisfy by default once a model is built specifically to be indistinguishable from a human on a video call ^[2]. For now, Tavus is treating that tension as a reason to go slowly: Griffin-Lite is not available to customers, only to select trusted testers as a research preview, while the company works through what it describes as safe disclosure features before any wider release ^[1][2].

The reaction hasn't been uniformly convinced. On Reddit, some commenters pushed back that Griffin's voice and responses felt too polished and rehearsed to be mistaken for a real person. Others raised a different objection entirely: one commenter argued that the computing effort spent on conversational polish would be better directed at what they called "functional intelligence" problems, rather than perfecting an avatar's human-likeness. On YouTube, an independent tester who got early access to multiple Griffin personas described the gesture and reaction accuracy as surprisingly good, but said it didn't feel human consistently, with occasional uncanny-valley moments breaking the illusion. The gap between the clean, controlled conditions of a benchmark study and an open-ended real conversation is likely where most of that skepticism lives.