Tavus claims new Griffin AI video model fooled nearly half of call testers

The AI start-up Tavus claims its new conversational video model is so lifelike that nearly half of the people who tested it over a video call believed they were talking to a real human.

During a series of one-minute face-to-face video calls, 26 out of 54 participants—exactly 48 per cent—mistook the AI, named Griffin, for a real person. That represents a massive leap from the company's previous model setup, Phoenix 4.5, which fooled just 2.4 per cent of participants under the exact same test conditions.

How the Griffin model works

The technological shift lies in how Griffin processes information. Traditional conversational AI systems operate like a relay race. One model translates speech to text, a large language model decides what to say, a voice generator synthesises the audio, and an animation engine finally renders the video avatar. This sequential approach inevitably creates awkward pauses and a clunky, turn-taking dynamic that immediately gives the AI away.

Griffin, by contrast, processes video and audio feeds simultaneously. Tavus says the model watches the user's video and listens in real-time while it is speaking. This allows the AI to behave far more naturally. It can interrupt, react instantly if a user cuts it off mid-sentence, and alter its facial expressions, gaze, and vocal tone on the fly based on what it sees and hears.

Significant caveats and security fears

However, the headline 48 per cent figure comes with significant caveats. The trial was highly limited, involving only 54 participants, and was conducted entirely by Tavus itself without independent verification. The company has not yet released the model widely, choosing instead to limit access to a version called Griffin-Lite while it builds out safety features and disclosure tools.

The rapid closing of the gap between AI avatars and human reality has already sparked intense debate. While some tech enthusiasts praised the breakthrough, others warned of the profound security risks. With video calls now standard for remote work, hiring, and sensitive corporate operations, the ability to effortlessly mimic human interaction raises serious identity-verification challenges.

Some online commentators suggested that lifelike video avatars of this quality should be illegal, warning of the potential for highly convincing video scams. Emad Mostaque, the former chief executive of Stability AI, reacted to the demo by declaring on social media that "remote work is cooked."