Tavus's AI Avatar Fooled 48% of People on a Live Video Call

Fifty-four people sat down for a one-minute video call with a stranger. Twenty-six of them walked away convinced they had just talked to a real person. They hadn't. The stranger was software.

That is the headline number from Tavus, a San Francisco AI startup, which announced its new "Human Interaction Model" Griffin on October 1 and says it just passed a real-time, video version of the Turing test. The New York Post picked it up this week, and the claim is now bouncing around tech Twitter: 48% fooled, up from 2.4% for Tavus's own previous stack. One founder of Stability AI, Emad Mostaque, summed up the reaction in two words: "remote work is cooked."

It is a hell of a number. It is also, on inspection, a number doing a great deal of marketing work — and the part of the story nobody is writing is what happens when this technology meets the people already faking their way through job interviews for a living.

Photo: Alan Turing at Princeton in 1936. His 1950 "imitation game" is the test Tavus says Griffin just passed — the company's framing, not an independent finding. Public domain via Wikimedia Commons.

The experiment, exactly as Tavus ran it

The details matter here because Tavus published them, and they are more modest than the headlines. The company recruited 54 participants through what it calls an "independent research platform" — the platform is unnamed, and no demographics were released — and told each of them they would be matched with another participant for a one-minute call about what they were looking forward to this year. Only after the call were they asked whether it had crossed their mind that their partner might not be a person. Then everyone was told the truth.

Twenty-six said they had been talking to a real human. And they were confident about it: those who believed it was human rated their confidence at 79% on average, versus 81% among those who guessed AI. People who caught it usually did so within the first 20 seconds. On seven-point ratings, participants scored the interaction 5.4 for naturalness, 5.6 for trustworthiness, 5.8 for willingness to talk again — and 4.9 for "the conversation flowed," the lowest score, which is where the seams still show.

The control is the awkward part. Tavus's previous pipeline — its Phoenix 4.5 renderer, Sparrow-2 timing model and Raven-1 perception model — fooled 1 of 41 people, or 2.4%. One person. That is a control group balanced on a single participant, and it is the entire baseline against which Griffin's "leap" is measured. It is shoddy statistics wearing a lab coat, and it tells you this was a product announcement dressed as science, not the other way around.

It also isn't the Turing test as Turing designed it. In the 1950 imitation game, the judge knows one of the two interlocutors is a machine and gets an extended session to probe. Tavus's participants were never told an AI might be present, so they had no reason to look for one; they got 60 seconds of small talk about their hopes for the year. Fooling an unsuspecting stranger for a minute is a real technical achievement. It is not passing the Turing test in any sense Turing would recognise.

What Griffin actually does under the hood

Strip away the test framing and the engineering is genuinely impressive. Griffin generates every pixel of every frame in real time from a single reference image — not just the face, but arms, fingers, the movement of the chair, the shadows it casts, the background. It runs at 720p in 320-millisecond chunks, with average audio-to-video latency of 0.43 seconds on Nvidia H100s, which Tavus says is roughly half the next published streaming-diffusion baseline. A voice clone needs about 10 seconds of reference audio. Conversational decisions — when to speak, hold, yield, interrupt, backchannel — happen at sub-second "mini-turns," and the avatar can be interrupted without losing its train of thought.

"Humans are evolutionarily designed to communicate face to face," CEO and co-founder Hassaan Raza wrote in the announcement. "We speak as much through our words as we do our expressions, tone, gestures and timing. So much of human communication is non-verbal. It's an art, a dance."

He went further on X: "A tutor with their eyes closed misses the confusion on your face. And an apology means very little when it comes with a smile. These details are the difference between feeling understood and having to manage the machine." Previous avatar systems, Raza noted, had a nasty habit of smiling through bad news. Griffin doesn't.

There is independent signal, too. Nvidia scored a research preview of the model, Griffin-Lite, on its VideoFDB benchmark in September: 3.83 on generation quality against 3.92 for a human reference and 2.80 for the next system tested, a Gemini 2.5 plus Anam cascade. On perception it scored 3.73 against the human's 4.20. Tavus's marketing line — "37% ahead" — is just the 1.03-point gap over that one rival cascade expressed as a percentage. A real lead over a real competitor, but a relative margin against a single setup, not an industry-wide coronation.

The company behind the avatar

Tavus was founded in 2020 in San Francisco by Raza and co-founder Quinn Favret, went through Y Combinator's summer 2021 batch, and started out making personalised AI video templates — record one video, clone your face and voice, generate thousands of variants for sales teams. Sequoia led a $6.1 million seed in March 2023. Then, last November, CRV led a $40 million Series B as the company launched PALs — "Personal Affective Links," its term for emotionally aware AI humans — with Scale Venture Partners, Sequoia, Y Combinator, HubSpot Ventures and Flex Capital participating.

"We've spent decades forcing humans to learn to speak the language of machines," Raza said at the time. "With PALs, we're finally teaching machines to think like humans — to see, hear, respond, and look like we do."

Tavus says it has raised $70 million in total, names Amazon, the Mayo Clinic and Salesforce as customers, and claims 150,000 developers and creators use its models — all company figures. Griffin-Lite, though, is not a product yet: it is available only to select trusted testers as a research preview, with customer-facing avatars still running the older Phoenix, Raven and Sparrow stack. No pricing, no general-availability date, no name for the bigger model Tavus says is coming.

Photo: The setting for millions of video interviews in 2026. Griffin's avatars are pitched as the face on the other side of the screen — currently in a closed research preview, not a shipping product. Photo by Nenad Stojković, CC BY 2.0, via Wikimedia Commons.

The part nobody is writing: the fraud angle

Here is the uncomfortable reading of Griffin. Everything Tavus celebrates — a face that breathes and blinks, interruptions handled naturally, confidence on both sides of the call — maps exactly onto the toolkit of the people already running fake-interview scams at industrial scale.

The FBI's Internet Crime Complaint Center warned back in 2022 about deepfakes in remote IT job interviews. Since then the problem has professionalised. On July 31 this year, the US State Department and the FBI, alongside partner countries, issued a joint alert that North Korea deploys IT workers under false identities to fund its weapons programmes — eight individuals were sentenced to prison in 2026 alone for running these schemes. In September, the FBI, Japan's National Police Agency, Australia's signals directorate and German intelligence jointly named the North Korean group WaterPlum, or "Contagious Interview": between December 2025 and July 2026 it compromised more than 30,000 devices and drained over 7,000 crypto wallets, moving roughly $10.71 million to North Korea, using fake job interviews as the lure. GetReal Security reports that 41% of IT, cybersecurity, risk and fraud leaders believe their company has hired and onboarded a fraudulent candidate. The FBI is separately investigating a scheme in which more than 100 companies unknowingly hired remote IT workers using AI-generated synthetic identities — fabricated résumés, cloned faces and voices in video interviews.

None of that is Griffin's fault. Tavus did not build a scam tool; it built a conversation tool, and it says it is developing "safe disclosure features" and working with AI-safety organisations before any wider release. But note what we found when we read the research post ourselves: those disclosure features get exactly two sentences. No partner names, no dates, no technical details. For a company claiming to have crossed the video Turing threshold, the safety plan is a paragraph, and the fraud industry it will inevitably serve is a billion-dollar machine with a two-year head start. The money story is instructive: a reported $108 million voice-clone fraud against Italy's Fideuram chairman in February, and the $25 million that Hong Kong's Arup wired after a deepfake video call with a fake CFO in 2024. Those were crude tools. Griffin is not crude.

Our verdict: a real milestone, a rented test, a coming disclosure fight

So where does that leave the 48%? It is a genuine engineering milestone wrapped in a rented scientific ritual. The rendering is real, the latency numbers are real, the Nvidia scores are real, and the jump from the company's own previous stack is real even if the baseline is one person wide. But a one-minute call with an unsuspecting participant who was never told to look for a machine is a demo with a confidence interval — roughly 35 to 61%, for the statistically inclined — not a verdict on machine intelligence.

The interesting question is not whether Griffin "passed" anything. It is who gets told they are talking to it. Tavus is right that the difference between feeling understood and managing the machine lives in the micro-details of a face. The corollary is that the difference between a product and a weapon lives in a disclosure banner. Some employers already ban candidates from using generative AI during live interviews; none have rules yet for when the interviewer is the AI. That rulebook is now being written, in a hurry, by companies that move faster than regulators. Whether Griffin ends up as the tutor that notices your confusion or the stranger that fooled 26 people in a minute will be decided less by the model than by the two sentences about disclosure Tavus has not yet written.

Related coverage: the safety benchmark that crashed AI safety scores from 97 to 3 with one free edit; OpenAI's fix that ended ChatGPT's delusions — after 7 suicides; and the night Claude routed a teen to cliff helicopters.

Sources: New York Post (October 10, 2026); Tavus official research announcement, tavus.io/griffin (October 1, 2026); Business Wire/Tavus press release (October 1, 2026); Decrypt; the-decoder.com; explainx.ai; digit.in; TechCrunch (Tavus funding history, 2023); HR Daily Advisor; startupfortune.com; IT Brew/FBI IC3 reporting. Quotes from Hassaan Raza are from Tavus's own announcement materials; the Mostaque reaction is via social-media reporting. Company figures (funding total, customer names, developer count) are Tavus's own claims.

Daniel Rao is a technology journalist covering AI, privacy and regulation at TechPulse USA. He writes about the collision between new technology and old norms.