Tavus’s Griffin avatar had 48% of people on a one-minute video call believe it was a real human — and the team called it the first to pass the “video Turing test.” It’s a genuinely impressive model. The point here isn’t to take the startup down a peg — it’s that the claim is bigger than the data, and what the Turing test measures was never what the headline implies.
The announcement, from Tavus: Griffin, a real-time video-to-video avatar, is “the first model to pass the video Turing test,” with 48% of people on a live one-minute call believing it was a real human — versus under 3% for previous systems. That’s a real leap in avatar realism, and the team deserves the credit. It raced around X, and most of the coverage amplified the 48%. Fewer people read the context — and the context is where it gets interesting.
Tavus designed, ran and reported it — on its own avatar — as most launch demos are. That’s normal for a product unveil; it just isn’t the same as an independent lab replicating the result.
Participants were told they were meeting “another participant,” not warned that one might be a machine and asked to hunt for it — which is the harder condition a formal Turing test uses. Suspicion changes behaviour a lot.
54 people, one minute each. A strong signal for a young product — not yet evidence that “half of people can’t tell.”
It means “about half didn’t suspect an AI during a short, friendly chat when nobody told them to look” — striking, but not “most people can’t tell.”
A fairer way to say it: “about half of 54 people didn’t suspect our avatar in a one-minute chat when we didn’t tell them to look.” That’s a genuinely impressive product result — it just isn’t “AI passed the Turing test.”
Here’s the part the whole debate keeps missing. The Turing test was never a test of intelligence. Turing called it “the imitation game,” and it scores exactly one thing: can a machine convincingly seem human? That’s a skill of mimicry and persuasion — not of reasoning, reliability, or judgment. In fact the two most famous ways to “pass” it are to make the pretense easier: OpenAI’s GPT-4.5 only cleared the text version when it was told to adopt a human persona and act dumber than it is; Griffin cleared its bar with a setup where no one was hunting for the machine.
Which should give you pause, not comfort. A system optimized to seem human is a system optimized to deceive one. “Passing the Turing test” isn’t a milestone on the road to general intelligence — it’s a milestone in persuasive imitation, and the better AI gets at it, the harder it becomes to know when you’re being fooled.
The bigger picture. Griffin is one data point in a wave of “surely this is AGI” moments — from AI proving Millennium-Prize-class theorems to passing the Turing test. We pulled the whole thread apart: why those signals are real, why neither is AGI, and the one question that actually tells you how close we are. → How close are we to AGI? The theorem-proving wave, decoded
Griffin is a real advance in real-time video avatars, and “indistinguishable from human on a short call” is a genuinely big deal — for fraud, for trust, for every call centre and video feed you’ll ever see. Take that seriously.
But “passed the video Turing test” is a marketing milestone, not a scientific one — a vendor’s own 54-person minute-long study, scored below half, with the hardest part of the real test quietly removed. And the finish line it implies — human-level machine intelligence — was never what the Turing test measured in the first place.