Tavus' Griffin: A one-minute conversation convinced nearly half of participants they were talking to a human
The San Francisco lab Tavus launched the video avatar Griffin on October 1, 2026, claiming it is the first model to pass a "video Turing test."
The San Francisco lab Tavus launched the video avatar Griffin on October 1, 2026, and claimed it is the first model to pass a "video Turing test." The basis: a study run by the company itself, in which 26 of 54 participants (48%) believed, after a one-minute video call, that they had spoken with a real human. The claim has drawn major attention — and immediate pushback from researchers, a Community Note on X, and criticism from former congressman Tom Malinowski. For now, access is limited to a Griffin-Lite research preview, with no full public rollout.
What the company claims
Tavus describes Griffin as the "world's first Human Interaction Model," a designation of the company's own invention. The announcement comes from CEO Hassaan Raza and chief research officer Ioannis Patras, according to Cryptobriefing.
The number that has stuck in coverage is 48 percent: in Tavus's own study, 26 of 54 participants believed Griffin-Lite was a human after one minute. By comparison, the company's previous system, Phoenix-4.5, under the same protocol convinced only 1 of 41 participants — 2.4 percent, according to UkrMedia. The leap is striking, but all the documentation comes from the company, and the result has not been independently verified.
How Griffin works
Technically, Tavus describes Griffin as a full-duplex video-to-video model. That means the model keeps watching and listening to the user while it responds itself — unlike turn-based systems that wait until the other party has finished. This enables interruptions, pauses, and immediate reaction when one party cuts the other off, which is central to natural conversational behavior.
According to Cryptobriefing, Griffin is built on Tavus's existing PAL platform, which the company says is used by more than 150,000 developers and businesses — also a company claim. The company further says Griffin merges the Phoenix and Raven models, and that it beat the next-best published system by 37 percent on real-time reaction metrics.
The benchmark numbers
Beyond the study, Tavus points to NVIDIA's VideoFDB benchmark, which is built on 237 clips from real video calls and covers 11 types of conversational behavior. The results as reported:
- Perception: Griffin scored 3.73 out of 5, against a human reference pair at 4.20.
- Audiovisual response generation: Griffin scored 3.83, versus 3.92 for real humans — while the next-best system on the list scored only 2.80, according to Dainik Jagran/MSN.
The distance to the human on the generation test (0.09 points) and to the next-best system (over one point) is the core argument in Tavus's framing of Griffin as a qualitative leap. But these numbers, too, are reported through the company's own statements and secondary coverage — there is no public technical report that can be checked.
Why the documentation is weak
PCMag, which covered the story on October 8, 2026, points to several problems. The result is "not yet independently verified," and the video calls were only one minute long — participants had little time to detect that they were talking to an AI-generated person. There is also no recognized standard for a "video Turing test"; the definition of what counts as "passing" is Tavus's own. The demo video on X received a Community Note questioning the claim.
PCMag also notes demo features such as breathing and eye and mouth movements, which point toward artifacts — a more cautious description than the more uncritical retellings claiming the demo was flawless.
The pushback: demands for mandatory disclosure
The rollout has triggered swift criticism. Former Democratic congressman from New Jersey Tom Malinowski wrote, according to PCMag:
"It should be illegal — plain and simple — for any AI to impersonate a human being without a clear and visible notice that it is in fact AI."
Researchers have similarly argued that the technology is unethical without clear AI disclosure, citing risks of fraud and impersonation — risks that become more concrete when a video-call avatar is perceived as human by nearly half of participants, even in short sessions.
Tavus's response
In a statement on X, as reported by PCMag and therefore marked as a company statement, Raza wrote: "We are working with partners on safeguards and disclosure systems, and inviting discussions with governments around broader regulation and safe use." He added that "people will always know they are interacting with AI," while the system provides an interface that removes the need to "speak computer."
Tavus has, according to PCMag, planned use cases in education and healthcare, for work with "low access or high cost." The company has delayed a full launch until the safety work is complete.
What remains unresolved
Several central questions remain open. There is no public technical documentation or study publication that would allow independent verification of the study design, participant recruitment, or benchmark results. The timeline for safeguards and disclosure systems is unspecified, as is the timing of a full public rollout — only the Griffin-Lite research preview is available so far.
For now, what is documented: 48 percent in a small, company-run study with one-minute calls, benchmark numbers that cannot be independently verified, and an ongoing debate over whether undisclosed AI impersonation in video calls should be legal at all.