Tavus Griffin Passes the Video Turing Test: When You Can't Tell AI From Human on a Video Call

Is it possible to distinguish between an AI and a human person during a video call? In a recent survey, the response was "no" for almost half of the participants. Griffin is a "Human Interaction Model" (HIM) that the San Francisco-based research lab Tavus says is the first AI to pass a real-time video Turing test. For anyone who uses video conversations for recruiting, business, or identity verification, the results are startling and extremely worrisome.

What Makes Griffin Different

Griffin is intended to replicate every aspect of human contact, in contrast to previous AI avatars that have a robotic or eerie feel. It can be seen laughing, showing surprise, interrupting, and responding to interruptions in a natural way. Even when it moves, waving, caressing its face, or tucking hair behind an ear, its visual portrayal stays steady and doesn't exhibit the distortion that usually distinguishes AI-generated footage.

54 participants in Tavus's live blind study had a one-minute video conversation with someone they thought was another study participant. It was actually Griffin. Twenty-six of them, or 48%, thought they had spoken to a genuine person when they ended the contact. In contrast, just 1 in 41 individuals were tricked by Tavus's prior model.

The Security Implications Are Profound

This innovation comes at a time when companies are already battling an increase in fraud enabled by deepfakes. Complex CEO fraud schemes, distant onboarding systems, and liveness detection are all being circumvented through real-time video modification. In one scenario that has been published, attackers authorized $25.6 million in transfers from a multinational company using a multi-participant video chat that was filled with deepfakes.

Griffin's skills elevate this danger. Visual verification becomes less trustworthy when a model can pass for a human in real-time conversation, complete with natural interruptions and emotional expressions. Forty-eight percent is essentially a coin flip, as one X user noted. After this, disclosure is no longer optional because the only trustworthy way to know you're speaking to an AI is if it tells you.

Why Tavus Is Holding Back

To its credit, Griffin has not yet been extensively released by Tavus. The model is only accessible as a restricted study preview, and the corporation claims that before implementing it more widely, it must figure out how to reveal its nature. That choice indicates an increasing awareness that, until protections catch up, the technology's potential for deception may exceed its advantages.

What This Means for Your Organization

The days of relying solely on video calls to verify identity are coming to an end. Organizations must implement multi-factor authentication, out-of-band confirmation, and recorded approval protocols as layers of verification that are independent of any one channel. Video conferencing solutions require integrated AI detection technologies, and security awareness training must change to address AI-generated video.

At Bayon Technologies Group, we assist businesses in strengthening their defenses against identity-based risks, such as fraud enabled by deepfakes. We make sure your staff can identify the fakes—and safeguard your bottom line—by providing everything from security awareness training and verification procedure design to incident response planning. Get in touch with us right now to find out more.