Tavus introduces Griffin, a model that reportedly passed a video Turing test
Tavus says its Griffin full-duplex video-to-video model made nearly half of testers believe they were speaking with a human during one-minute video calls.
Startup Tavus introduced Griffin, a model that, according to the company, passed a video version of the Turing test.
In a live one-minute video call, 48% of participants (26 out of 54) believed they were speaking with a real person rather than an AI. The demos look strikingly lifelike and have already generated buzz on social media, with Tavus's announcement reaching 10 million views.
Griffin laughs, changes tone, interrupts, can be interrupted, and senses pauses well. It also follows instructions closely and understands the interlocutor and their environment — for example, it can help solve a Rubik's cube. That is before mentioning smaller human-like behaviors, such as spinning in a chair or smiling like a real person.
Griffin works as a unified full-duplex video-to-video system. Every fraction of a second it decides whether to stay silent, nod, agree, or take over the conversation, while continuing to watch and listen even when it is speaking. On top of this manager, streaming speech and video generators turn those signals into a voice and a face in real time. Some architectural details are described in the blog post.
The model is currently available only to selected testers as a research preview, because Tavus is confident that Griffin can deceive a person into thinking they are not talking to an AI. The company promises a release very soon after working on safety and developing watermarks.
Why it matters
AnalysisIt shows how close real-time AI video agents are getting to human-like interaction and why disclosure, safety, and watermarking matter before broad deployment.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
The Turing Test and Its Obsolescence
The historical artificial intelligence test proposed by Alan Turing in 1950 assesses a machine's ability to engage in text dialogue convincingly enough that a human cannot distinguish it from another human. This entry explains why the advent of language models has rendered this test scientifically irrelevant.
Sora and AI Video (Video Generation from Text)
A cutting-edge class of generative neural networks capable of creating photorealistic dynamic videos from text or static images (OpenAI Sora, Runway Gen-3, Kling, Luma Dream Machine).
14 GitHub Repositories for Making Video in Claude Code or Codex
A curated list of GitHub tools for breaking down references, building motion graphics, adding 3D, and rendering finished videos with code and AI agents.
Complete Guide to Higgsfield Cinema Studio 4.0: Scene Management, Soul 2.0, and API Automation
In-depth practical breakdown of the major Higgsfield update: Cinema Studio 4.0, consistent characters with Soul 2.0, API-based generation, and a ready-to-use UGC pipeline.