Tavus Previews Video-Call AI That Passed for Human in 48% of a Small Company Test

In a company study, 26 of 54 participants mistook Tavus’s Griffin-Lite for a person after a one-minute video call. Tavus previewed Griffin on October 1, 2026, but is keeping the lighter version in a research preview for trusted testers. That finding captures a convincing first impression—not yet evidence that the illusion holds through a longer conversation.

A face that reacts while you speak

Griffin is designed for live, face-to-face exchanges rather than a prerecorded talking avatar. Tavus calls it a Human Interaction Model: a system that processes speech, facial expressions, tone, gestures, and pauses while receiving and generating video. The company says it can watch, listen, speak, and gesture at the same time.

Those overlapping behaviors are central to the pitch. Tavus describes Griffin nodding while someone speaks and incorporating visual details into its reply. The intended effect is a continuous exchange, where the avatar responds to more than the words in a question. That is a different target from simply producing a realistic face or keeping lips aligned with speech.

The test began with an expectation of a human

The study’s setup is essential to interpreting the headline result. Participants expected to meet another participant, rather than entering the call knowing they would encounter AI. They were debriefed afterward.

Tavus reported a much lower figure for its previous system: one of 41 participants, or 2.4%, mistook it for a person. The comparison suggests a substantial change in the company’s short-call results, but the samples are small.

Looking human and reading people are different scores

Tavus also points to NVIDIA’s VideoFDB benchmark, which separates generated conversational behavior from understanding conversational cues. Its scores provide another view of the model, but they are not another count of people fooled by an avatar.

• Generation: Griffin-Lite scored 3.83 out of 5, close to the 3.92 for human reference recordings. Gemini 2.5 plus Anam, the next listed system, scored 2.80.
• Perception: Griffin-Lite scored 3.73 against 4.20 for human reference recordings, leaving a wider gap in understanding conversational cues.

The near-human generation score has limits. The table includes only two competing AI systems, and a language model scored nonverbal behavior against set criteria; human reference recordings served as a comparison point. This should not be read as a human jury finding Griffin equally capable across conversation.

Access waits on safety and disclosure

Tavus lists tutoring, rehearsing difficult conversations, and camera-based technical support as potential uses. These are proposed applications, not outcomes established by the one-minute study. A persuasive first encounter and useful assistance are separate questions; the preview’s reported tests focus on human-likeness and conversational behavior.

Tavus is working on safety and disclosure features before wider access, but has given no firm public launch date. The company says a more capable version will follow once safety concerns are addressed. How it will enforce AI disclosure and check consent to copy someone’s identity remains unresolved—a concrete rollout question for a system whose appeal includes being mistaken for a person.