AI video chatbot fools 26 out of 54 testers into thinking it's human

During live video calls, Tavus' Griffin AI managed to convince 26 out of 54 participants they had spoken to an actual person. The company is withholding the model from commercial use precisely because of that success.

According to Tavus, Griffin is the first AI to successfully pass the video Turing test. In a comparable test, the firm's earlier system only deceived one out of 41 test subjects.

What Enabled Griffin to Fool 26 Strangers?

Hassaan Raza, Tavus co-founder and CEO, said that human communication relies equally on expression, tone, and timing, not just words. He noted that machines typically force users to adapt by providing commands, clicks, or prompts.

The Tavus CEO is promoting a computer that interprets tone and expression as a human would, and also has a human-like appearance.

Introducing Griffin, the first model to pass the video Turing test.

48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.

It’s the first Human Interaction Model (HIM). pic.twitter.com/eb11XbC8Xr

— Tavus (@tavus) October 1, 2026

Tavus compares Griffin to cascade systems, which sequentially run speech recognition, language models, and voice synthesis. Those systems hold their reply until the user has finished speaking.

Griffin reassesses the dialogue at sub-second intervals, simultaneously interpreting words and nonverbal signals. It also processes video input, something Tavus claims most interactive AI systems ignore.

This architecture allows Griffin to nod and voice acknowledgments such as 'mm-hm' while a person is speaking, and it can resume its point after being interrupted.

“Griffin-Lite’s speech generation produces high-quality speech at conversational latency and can clone a speaker’s voice from about 10 seconds of audio,” Tavus claims.

The VideoFDB benchmark from NVIDIA evaluates an AI's ability to conduct live video conversations. Griffin received a score of 3.83 out of 5 on this benchmark.

Human reference recordings achieved 3.92, whereas the top competing system — a combination of Gemini 2.5 and Anam — scored only 2.80.

Why Is Tavus Withholding Its Griffin AI from Customers?

Test subjects anticipated a one-minute conversation with a human about their plans for the upcoming year. Over half admitted that the possibility of an AI interlocutor never occurred to them.

Any doubts that arose typically emerged within the initial 20 seconds. Tavus recognizes that the same lifelike behavior could trick people into thinking they are interacting with a human.

Consequently, the preview version called Griffin-Lite is accessible only to a few trusted testers. The firm states it is developing disclosure features and working with AI safety groups prior to a broader rollout.

The threat is not theoretical, as demonstrated by cryptocurrency fraud. TRM Labs reported that criminal AI adoption increased by 40% year-over-year in 2026, driven primarily by scams.

Should models similar to Griffin proliferate, a recognizable face on a video call might offer little assurance about the caller's identity.