According to Tavus, its new model, Griffin, persuaded almost half of those who engaged with it during a video call that they were speaking with an actual person. The company claims that among the participants who took part in the test, 26 out of 54 thought Griffin-Lite was human following a one-minute conversation, compared to only one out of 41 who felt the same way about the older system.
Tavus’ own research page is where these findings come from. Participants were told they would be paired with someone else for a one-minute call about what they were anticipating this year. The question of whether their partner might not be real came only at the close of the experiment.
The Numbers Behind Griffin-Lite
Tavus says Griffin-Lite ranks first on Nvidia’s VideoFDB benchmark. On the generation track, which grades how natural and expressive a model’s responses are, Griffin-Lite scored 3.83. The next-best system got 2.80, and a human reference scored 3.92.
This track tests whether a model grasps what it perceives through sight and sound, revealing where it falls short of human performance. The strongest baseline came from Griffin-Lite, which scored 3.73, compared with 3.44, while the human reference reached 4.20. Nvidia conducted the evaluation on its own.
Why This Matters Outside Tech
North Korea-linked hackers have already used video calls for deception. In January, they employed deepfakes, which are videos made by AI that mimic a real person, during Zoom or Teams calls to impersonate trusted contacts. The intrusion was attributed to BlueNoroff, a subsidiary of the Lazarus Group.
People are tricked into putting malware on their devices by being told it fixes audio problems. According to David Liberman, who helped create Gonka, a system for shared AI processing, photos and videos can no longer be trusted as evidence that something is true.
What Griffin Can Actually Do
Tavus’ Griffin operates with two-way traffic: it hears, observes and speaks all at once, like a telephone conversation rather than a walkie-talkie exchange. During a demonstration, it guides a person through solving a Rubik’s cube by watching his hands, pausing when he falls silent to think.
The delay between audio and video on NVIDIA H100 chips, the type used in AI data centers, runs an average of 0.43 seconds, according to Tavus, which puts it at half the time of the next quickest approach.
The Test Was Self-Conducted
Tavus ran the study on its own. It brought in participants through what it describes as an independent research platform. A community note on X has already pointed out that the findings lack independent verification and do not follow a standard protocol.
According to Tavus, those who grew suspicious tended to do so after about 20 seconds.
Griffin-Lite Is Locked Down
The Griffin-Lite service is not open to regular customers. Only a small group of trusted testers can ask for permission by filling out a form on the Tavus website. The firm has said it is developing disclosure features and collaborating with AI safety organizations.
Tavus raised a $40 million Series B in November 2025, led by CRV. The system that scored 2.4% stitched together three separate models, one each for visuals, dialogue and perception.
Our Take on the Claim
Tavus has announced a milestone, though the research comes from its own labs and has not been verified by anyone else. The firm contends that Griffin requires safety measures prior to a public launch, and the trusted-tester arrangement means no one outside that select group can attempt it for themselves.
Shipping safely is a sensible goal for any firm, yet that approach also means the assertion depends on a study that has not been examined by anyone outside Tavus.
Key Facts Box
- Belief rate: 26 out of 54 participants (48%) believed Griffin-Lite was human
- Old system belief rate: 1 out of 41
- VideoFDB generation score: Griffin-Lite 3.83, next-best 2.80, human reference 3.92
- VideoFDB perception score: Griffin-Lite 3.73, strongest baseline 3.44, human reference 4.20
- Series B funding: $40 million, led by CRV
- Audio-to-video delay: 0.43 seconds on NVIDIA H100 chips
Source material: “This AI Is Already Fooling People on Video Calls Into Thinking It's Human, Company Says,” Decrypt.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.