Tavus launches Griffin: video-conversation AI that reportedly fooled 48 percent of test participants
AI company Tavus launched Griffin on October 1, 2026, a model that conducts real-time face-to-face video conversations. The company calls it the first of its kind and says that 48 percent of participants in its own study believed they were speaking with a human. But the "video Turing test" claim has already been flagged with a community note on X stating that the results are not independently verified, and technology journalists are warning of fraud and harm risks — while the model remains open only to a small group of testers.
What Tavus announced
On October 1, 2026, Tavus, which describes itself as a "human computing" company, announced Griffin in a paid press release via Business Wire. The company describes Griffin as the "first Human Interaction Model" — a self-applied label, not an established category — and as a full-duplex, video-to-video model that conducts face-to-face conversations in real time. An X post about the announcement has reportedly drawn more than 13 million views.
"For decades, we've imagined computers as partners, not just tools," said Tavus CEO Hassaan Raza in the press release.
In practical terms, full-duplex and video-to-video means the model both sees and generates video — a moving, talking face — while listening and responding, rather than running speech-to-text, text generation, and speech-to-speech as separate steps. This, the company says, is what enables a conversation that feels like sitting across from a person.
The numbers Tavus itself presented
All key figures come from Tavus and have not been independently confirmed:
- On NVIDIA's VideoFDB benchmark, Griffin is said to have scored 3.83 out of 5 on generation, versus a human reference of 3.92 and 2.80 for the most-published competing system.
- In the company's own live study, 48 percent of participants reportedly believed they were speaking with a real human rather than an AI.
- The company's previous system, tested the same way, fooled only 1 of 41 participants, according to Tavus.
The company itself concludes that Griffin is "the first to pass the video Turing test." It is precisely this formulation that has drawn pushback.
The contested test
According to Protos, Tavus's Turing-test claim was supplied with a so-called "community note" on X — a user-generated context label — shortly after the announcement, stating that the results are "not independently verified" and not based on a "standard protocol."
There is also no published, peer-reviewed description of the study design. Protos reports that the study involved 54 participants recruited via an independent research platform, and that 26 of them believed they were speaking with a human. Here it is worth being cautious: Protos's own text contains an internal inconsistency, with a phrasing of "out of 26 people, 48 percent," which cannot be reconciled with the 26-of-54 frame (which corresponds to 48.1 percent). The precise participant count therefore cannot be determined from the available coverage, and the denominator should be treated with care.
Another detail that cuts against the strongest reading of the result: according to Protos, most suspicion arose within the first 20 seconds of the conversations. This could mean that the participants who were "fooled" did not necessarily hold that belief throughout the session — but the interpretation depends on methodology that has not been made public.
According to Protos, Tavus does not provide demographic data about the participants — neither age, technical background, nor how aware they were that they might encounter an AI. For a result intended to document human convincingness, the absence of such context is a significant limitation.
The reactions: fraud and psychological harm
The debate has not only been about methodology. According to Protos, technology journalist Kevin Roose has characterized what Tavus has built as, in effect, a "machine that tricks old people into handing over their bank passwords." Protos's coverage also cites concerns tied to AI-driven fraud scenarios and psychological harm — a realistic, talking video persona could in principle be used to impersonate known people or to exploit vulnerable users.
Tavus has itself acknowledged that it is precisely the degree of realism that makes safety work necessary before the technology is released to the general public. The company says it is working on so-called disclosure features — ways of revealing that the other party is an AI — and, according to its own statement, is collaborating with AI safety organizations ahead of a broader launch.
What actually exists today
The full Griffin model is not publicly available. The only thing that exists is Griffin-Lite — a preview version open only to a select group of "trusted testers." The company has not provided a launch date, price, or timeline for broader access.
That means all assessments of Griffin in practice rest on the company's own demonstrations and its own numbers. Neither the VideoFDB result, the study, nor the "video Turing test" claim has been confirmed by an independent party.
The open questions
Much remains before Tavus's claim can be judged:
- No independent evaluation. The community note on X points to the most fundamental issue: no one but Tavus has tested the system and reported the numbers.
- Unclear methodology. Recruitment via a research platform is all that is known about the sample; demographics, instructions, and scoring criteria have not been published, and there is no standard protocol against which a video Turing test can be measured.
- Unclear denominator. The discrepancy in the public rendering of the study (26 of 54 versus an "out of 26" phrasing) must be clarified by Tavus itself.
- No commercial or timeline plan. When Griffin will become broadly available, and on what terms, is unknown.
What can be established is that Tavus has announced a model of a kind no one else has marketed in the same way, that the company itself presents numbers placing it close to a human reference — and that the same realism meant to impress is the reason the model remains locked away.