AI Just Hit 48% Human on Video Calls — Blockchain Bros, Your KYC May Need an Upgrade

AI startup Tavus unveiled Griffin on October 1, a model it describes as the first Human Interaction Model, and said 48% of people who spoke with the system on a live one-minute video call believed they were talking to a real person. Griffin-Lite, the version tested, faced 54 participants, and 26 said afterward they believed their partner was human. The company's previous system scored 2.4% on the same test, having faced 41 people and convinced exactly one. Participants were told they would be matched with another person for a video call about what they were looking forward to this year; only at the end were they asked whether it had crossed their mind that their partner might not be real. Nobody on the call was told a bot might be on the other end. Tavus said the people who grew suspicious usually did so within 20 seconds.

The setup is not the classic Turing test proposed in 1950 by British mathematician Alan Turing, in which a judge speaks with a hidden human and a hidden machine and decides which is which. A community note on X has flagged that the results are not independently verified and do not follow a standard protocol. The data comes from Tavus' own research page, and participants were recruited through what the company calls an independent research platform. Griffin adds a real-time face and voice to a category where text-only systems have already shown strength, including OpenAI's GPT-4.5, which a UC San Diego study found convinced judges it was human in 73% of conversations when prompted to play an introverted, internet-savvy young person.

On NVIDIA's VideoFDB benchmark for live audio and video conversation, Tavus said Griffin-Lite ranks first. Its generation track, which grades how natural and expressive a model's responses are, gave Griffin-Lite 3.83, compared with 2.80 for the next-best system and 3.92 for the human reference. On the perception track, which checks whether a model understands what it sees and hears, Griffin-Lite scored 3.73 versus 3.44 for the strongest baseline, with the human reference at 4.20. NVIDIA ran the evaluation independently, according to Tavus.

Griffin operates in full duplex, listening, watching and talking at the same time rather than in turn-based exchanges. In a Tavus demo, the model coached a man through a Rubik's cube based on what it saw in his hands and waited when he went quiet to think. Audio-to-video delay averaged 0.43 seconds on NVIDIA H100 chips, which Tavus said is half that of the next fastest method. The company framed its work against a backdrop of documented abuse, noting that in January, North Korea-linked hackers used deepfake video on Zoom and Teams calls to pose as recruiters and other targets.