Tavus has introduced Griffin, a Human Interaction Model (HIM) designed to pass the video Turing test. In live video call trials, 26 out of 54 participants believed they were speaking to a human. This marks a significant improvement over the company's previous system, which successfully fooled only 1 out of 41 participants.
Unlike traditional cascade systems that process speech, language, and synthesis sequentially, Griffin evaluates conversations every fraction of a second. The model analyzes both verbal and nonverbal cues, including video input, allowing it to react with nods or conversational fillers while a user is still speaking. It maintains its place even when interrupted.
In the NVIDIA VideoFDB benchmark, Griffin achieved a score of 3.83 out of 5, compared to 3.92 for human reference recordings and 2.80 for the next-best system, which combined Gemini 2.5 and Anam. Additionally, the Griffin-Lite version can clone a speaker's voice using approximately 10 seconds of audio.
Tavus is currently restricting access to Griffin-Lite to select trusted testers. The company cites concerns regarding potential deception and misuse, noting that criminal AI adoption in scams rose 40% year-on-year in 2026. Tavus is developing disclosure features and collaborating with safety organizations before a broader release.
