Tavus Unveils AI Video Model That Fools Nearly Half of Test Callers

Artificial intelligence startup Tavus claims its newest video generation model can successfully impersonate a human during live conversations, fooling nearly half of the participants on test calls. The company unveiled the technology on October 1, branding it the first "Human Interaction Model" designed to navigate face-to-face dialogue.

During a test of the Griffin-Lite version, 54 participants engaged in a one-minute video call discussing what they looked forward to this year. Afterward, 26 of those individuals stated they believed they were speaking to a real person. An older Tavus system achieved just a 2.4% success rate on the same test, convincing only one out of 41 participants.

Participants in the trial were not informed that an artificial intelligence might be on the line, differing from the classical Turing test where judges actively try to distinguish human from machine. Tavus published these findings on its own research page after recruiting participants through an independent platform. However, a community note on X quickly flagged that the results lack independent verification and do not adhere to standardized testing protocols.

Technical Leap in Real-Time Interaction

Griffin represents a significant architectural shift from older systems that stitched together separate models for visuals, dialogue, and perception. The new model is full-duplex, meaning it listens, watches, and talks concurrently rather than operating like a walkie-talkie. This allows the system to pause when a user goes quiet or adapt dynamically to physical cues, such as coaching a user through a Rubik's cube based on visual input.

On NVIDIA's VideoFDB benchmark, which evaluates live audio and video conversations, Griffin-Lite ranked first in its generation track with a score of 3.83, trailing just behind a human reference score of 3.92. The model achieved an average audio-to-video delay of 0.43 seconds when running on NVIDIA H100 chips inside AI data centers.

Why it matters

While the technology showcases rapid advancements in generative AI, it arrives alongside growing apprehension regarding digital fraud. Security researchers have already documented instances of threat actors utilizing deepfake video on collaboration platforms like Zoom and Teams to pose as trusted contacts and deploy malware.

As conversational video models become more responsive and realistic, verifying the identity of participants on remote calls grows increasingly difficult. Companies have begun implementing spontaneous verification protocols, such as asking for specific local details or unexpected identification, to counter sophisticated social engineering tactics.

Tavus has restricted public access to Griffin-Lite, making it available only to select trusted testers through an application form. The company states it is actively developing disclosure features and collaborating with safety organizations before considering any broader release.

What to watch

Observers will monitor whether independent benchmarks can replicate Tavus's impressive latency and perception scores. Furthermore, industry watchers are tracking how platform developers and cybersecurity firms plan to build robust authentication defenses against real-time video impersonation.