Tavus says nearly half the people who spoke to its new Griffin model on a live video call thought it was a real person, up from fewer than 3% for earlier systems. The company is holding the model back while it works on safeguards, but the result shows how quickly the last easy way to check who you are talking to is disappearing.

[For more news, click here] 

The people in the study thought they had signed up for something ordinary, a one-minute video call with another participant to chat about their upcoming plans. The face on the other end listened, laughed, reacted when they held things up to the camera, and talked back. Only after the call, and after a short survey, were they asked whether they thought their partner had been real.

Twenty-six of the 54 people said yes, but their partner was an AI system called Griffin, built by the San Francisco startup Tavus, which announced the model on 1 October and called it the first to pass a video Turing test. When Tavus ran the same kind of test with its previous system, Phoenix-4.5, one person out of 41 was fooled.

Griffin also took first place on VideoFDB, a benchmark built by NVIDIA researchers to measure how well AI systems hold a live, two-way video conversation, scoring 3.83 out of 5 for the quality of its responses against 3.92 for real humans and 2.80 for the next-best system, according to Runtime Wire.

What a Video Turing Test Does and Does Not Prove

The numbers need some care, since Tavus ran the study itself, the sample was small, and the calls lasted only a minute. A longer conversation, or someone who suspected a trick, might have gone differently. Tavus also calls Griffin a Human Interaction Model, a category it named itself.

Alan Turing proposed his imitation game in 1950 as a test of whether a machine could pass as human in typed conversation, and in a 2025 study at the University of California, San Diego, people picked OpenAI’s GPT-4.5 as the human 73% of the time when it was told to adopt a persona. Cloned voices followed and video was the last place where most people still trusted their own eyes. For many families and businesses, “turn your camera on” has been the simplest way to check that someone is who they say they are.

When Seeing is no Longer Believing

In 2024, an employee at the engineering firm Arup in Hong Kong transferred about $25 million after a video call in which the company’s Chief Financial Officer and several colleagues were all deepfakes. In 2022, the FBI warned that people were using deepfaked video to interview for remote jobs. A 2025 study by the biometrics company iProov found that only 0.1% of 2,000 people in the US and UK could correctly tell every real image and video from every fake one they were shown.

Most of those fakes were stitched together ahead of time, whereas Griffin works live. It watches and listens while it talks, decides when to speak, wait, or let the other person cut in, and generates a whole scene rather than just a talking face, producing each response in under half a second on NVIDIA’s H100 chips. For the majority of people, that is quite similar to how a person behaves on a real video call and they may not believe it if they were told the person on the other line is not real.

What Tavus Says it is Doing About the Risk

“The same properties that make Human Interaction Models powerful interfaces for natural communications between human and machine allow them to deceive a human into believing it is not AI,” Tavus wrote on its Griffin page, an unusually blunt admission from a company launching a new model.

For now, a lighter version called Griffin-Lite is open only to selected testers as a research preview, and Tavus says it will not offer the model to customers until it has built disclosure features and other safety measures, working with AI safety organizations. The company says more than 150,000 developers and businesses use its tools, including Amazon, Mayo Clinic, and Salesforce, and it has raised $70 million from investors including Sequoia Capital and CRV, according to its announcement.

Its Chief Executive, Hassaan Raza, has described the goal in terms that sound very different once you know the test results. “At Tavus, we believe in a future where machines meet us where we are,” he wrote. “We want computing to become invisible.”

Living with Doubt

There are good uses for a video agent that feels natural to talk to, from patient check-ins to tutoring to customer support for people who struggle with menus and chat windows. The trouble is that the qualities that make it useful are the same ones a scammer wants.

The European Union’s AI Act requires that people be told when they are interacting with an AI system, and in the US the Federal Communications Commission ruled in 2024 that robocalls using AI-generated voices fall under existing bans. Neither was written with a live, interactive video face in mind, and enforcement depends on the people building these systems choosing to disclose them.

Until the rules catch up, some of the most practical defenses against a convincing fake on a video call are low-tech. Security teams already tell staff to confirm payment requests on a second channel, and some families have started agreeing on code words for emergencies. Tavus has not said when Griffin will be available to its customers.

Related Articles

An Exposed Server Revealed How a Ransomware Affiliate Used AI to Plan Attacks

Exclusive: Nikita Astionov on the Security Risks of Giving AI Agents More Control

Why AMD is Buying Fei-Fei Li’s World Labs for $8.2 Billion to Compete in Physical AI