AI Tricking Video Calls: Believing It’s Human, Company Says

This AI Is Already Fooling People on Video Calls Into Thinking It’s Human, Company Says

AI startup Tavus says its new conversational video model, Griffin, persuaded nearly half of participants in an internal test to believe they were speaking with a real person on a live video call.

According to Tavus, 26 of 54 people who interacted with Griffin during a one-minute video call concluded it was human. The company says people who became suspicious generally did so within about 20 seconds. Tavus has not released Griffin to retail customers yet.

The results come from Tavus’ own research page, and the company says participants were recruited through what it described as an independent research platform. A community note on X has flagged that the findings are not independently verified and do not follow a standard protocol.

Tavus also pointed to performance on NVIDIA’s VideoFDB benchmark, which evaluates live audio-and-video conversation systems. On the benchmark’s “generation track,” Tavus says Griffin-Lite scored 3.83 for naturalness and expressiveness. Tavus said the next-best system scored 2.80, while the human reference scored 3.92.

The claims underline how quickly real-time, humanlike video agents are improving—an important development for crypto and fintech, where video calls are increasingly used for customer support, onboarding, and identity checks.

At the same time, the same capabilities can amplify fraud risks. Security guidance and reporting in recent years has warned that deepfake video calls are becoming a practical tool for social engineering, allowing scammers to mimic familiar faces and voices to pressure targets into sending money or disclosing sensitive information.

Separate disclosures highlighted how AI-driven impersonation can be operationalized at scale. In one described setup, operators mixed AI personas with real gig workers in roughly a 3-to-1 AI-to-human ratio. The AI accounts handled routine chat, while humans stepped in for actions the bots could not perform—such as live video calls or social-media verification—helping reduce suspicion. The operation also tracked which users began to suspect they were talking to a bot.

Public warnings have increasingly focused on the financial impact of these tactics. For example, a widely reported case involved a worker transferring $25 million after being deceived by a video conference impersonation. Security experts have urged basic verification steps, including shared code words for sensitive requests.

For crypto platforms, exchanges, and payment apps, the trend puts more pressure on identity and account-security workflows. As lifelike video agents improve, organizations may need stronger controls beyond “seeing is believing,” including layered verification and policies that treat urgent payment instructions over video as a high-risk event.