The 54-Person Test Nobody Knew They Were Taking
Picture this. You join a quick one-minute video call with a stranger. You chat about what you are looking forward to this year. The person across the screen nods, reacts, interrupts you when you pause too long, and gets interrupted right back without losing their train of thought. You hang up feeling like you just met someone real.
Except you never did. That person never existed. Every pixel of them was generated in real time by an AI model called Griffin, built by a San Francisco startup named Tavus. And according to the company’s announcement, 48 percent of the people who took that call were fooled.
This is the moment the Turing test left the text box and walked onto a video call. The question is no longer whether a machine can sound human. It is whether you will even notice the difference when it is staring back at you through your webcam.
Tavus calls Griffin the first “Human Interaction Model.” The company rolled it out in an X post at the start of October, and the New York Post picked up the story this week, calling the whole thing creepy. Which, honestly, feels fair.
Here is how the test worked, based on the detailed breakdown from Tavus and the independent analysis published by ExplainX:
- 54 participants were recruited through an independent research platform.
- Each was told they would be matched with another participant for a one-minute call.
- The “participant” on the other end was actually a Griffin-powered avatar, generating face, voice, and replies live.
- After the call, participants were only then asked whether it had crossed their mind that the other person might not be human.
26 out of 54, a full 48 percent, said the person was real. The people who believed it were 79 percent confident on average. The ones who called it AI were 81 percent confident, and most of them spotted something wrong within the first 20 seconds.
For comparison, the company’s older system, a stack of three separate models handling vision, conversation, and perception, scored just 2.4 percent on the same protocol. Previous systems overall came in under 3 percent. So Griffin did not just improve on the state of the art. It jumped the fence.
What Griffin Actually Does
The technical trick here is what Tavus calls full-duplex AI video. Older AI avatars worked like a movie being assembled: script the words, render the face, then play it back. Griffin generates every pixel of every frame in real time, from a single reference image. That includes the face, the arms and fingers, the movement of the chair, the shadows it casts, even the background.
The avatar breathes. It blinks. It adjusts its explanation when it sees your facial expressions change. It can interrupt you, and it can be interrupted without derailing. In demos, Griffin taught a participant how to solve a Rubik’s Cube, reacting to visual cues in real time, and held its own in a game of Simon Says.
On NVIDIA’s VideoFDB benchmark for video generation, Griffin scored 3.83 out of 5. The human reference scored 3.92. The next-highest published system scored 2.80. That is the number that should make you sit up: we are within a tenth of a point of human, as measured by the company that makes the chips running the whole AI boom.
Tavus CEO and co-founder Hassaan Raza framed it philosophically: humans are evolutionarily designed to communicate face to face, he said, and so much of communication is non-verbal. An art, a dance. The company’s vision is a student facing an AI tutor and explaining what they did not understand, or a worker rehearsing a difficult conversation with a machine that reads the room back.
Nice vision. Terrifying possibilities. Let us talk about those.
The Scam Machine Argument
Tech journalist Kevin Roose did not mince words. He said the company has essentially built a machine for tricking older people into handing over their bank passwords. A Twitch streamer called Mako Fukasame put it even more bluntly: the only apparent purposes are to drain your grandparents’ savings and to replace you in a job interview.
They are not being hysterical. Video-call fraud is already one of the fastest-growing categories of financial crime, and it has been held back mostly by the fact that deepfake video was laggy, glitchy, and easy to spot in live conversation. Griffin removes exactly those friction points. The avatars come in male and female looks, they react in real time, and they eliminate the old reality-breaking tells like smiling when hearing bad news.
If 48 percent of alert, participating adults could not tell the difference in a controlled one-minute test, what happens when the person on the other end is your bank’s “fraud department” calling about an urgent transfer, or your “boss” asking you to approve a payment, or a relative who says they need money wired right now?
The company says the models can also be used for students in schools or as tutors. That may well be true. But the safety case has to answer the scam question first, because the scammers will not wait for the tutoring curriculum.
The Catch: Did It Really Pass the Turing Test?
Before we hand Griffin the trophy, there is a significant asterisk, and it is worth taking seriously. On X, Tavus’s original announcement was community-noted with a disclaimer: the results are not independently verified and do not follow a standard protocol. And that matters.
In the classic Turing test, proposed by Alan Turing in 1950, the judge knows they are talking to one hidden human and one hidden machine, and the game is to figure out which is which. In Tavus’s test, participants were told they were talking to another participant. They had no reason to suspect anything. Being fooled when you were never told to be suspicious is a very different feat from fooling an adversarial judge.
It was also a one-minute call about a light topic, run on Tavus’s own protocol, with 54 people. The lowest scoring axis, interestingly, was conversation flow, which got a 4.9 out of 7. Something felt slightly off even to the people who bought it.
None of that makes Griffin unimpressive. Going from 2.4 percent to 48 percent is a genuine leap. But it does mean the honest framing is this: a San Francisco startup’s internal study suggests its AI avatar can fool about half of unsuspecting people for about a minute on video. That is still a big deal. It is just not a coronation.
What This Means for the Rest of Us
The honest truth is that the technology is out of the lab now. Tavus raised a $40 million Series B round led by CRV last November, and this launch is clearly aimed at customers who want to put these avatars in front of real people: customer support, sales, education, recruiting. The job interview angle is the one getting all the headlines, and for good reason. If a machine can hold a convincing video conversation in real time, the line between “talking to the company” and “talking to the company’s avatar” is about to get very blurry.
A few practical takeaways. First, verification habits need to catch up with video. Voice codes and callback routines that worked for phone scams may need to become standard for video calls too, especially for anything involving money. Second, companies deploying this will need to be loudly transparent about it, because the first big scandal involving an undisclosed AI on a customer video call will poison the whole category. Third, regulators are still thinking in text. California is already pushing AI oversight rules, but video-call impersonation is moving faster than the rulebook.
The broader pattern is one we keep seeing across the technology beat this month. Google just reshuffled what its free AI tier can do, OpenAI’s revenue numbers rattled Wall Street, and now a startup is arguing that seeing is no longer believing. The AI race has entered its video era, and it is going to test every institution that relies on a face on a screen.
Here is the bottom line. Griffin may or may not have truly passed a Turing test. But it did something more important: it proved that the remaining gap between an AI avatar and a human on a video call is now small enough that half of us cannot find it in a minute. That gap is only closing. The question is whether our instincts, our laws, and our scam defenses can close the distance just as fast.
