Nearly half of test subjects mistook Tavus' AI video avatar for a real person on a one-minute call
Tavus has introduced a new AI system called Griffin, positioned as the first model purpose-built for real-time human interaction over video. Unlike earlier AI avatars that were largely one-directional or limited in responsiveness, Griffin processes multiple social signals at once - facial expressions, tone of voice, and physical gestures - and responds within a live video call context.
The benchmark that has drawn the most attention is from a Tavus-conducted study: 48 percent of participants believed they were speaking with a real person after a one-minute call with Griffin. That figure stands in sharp contrast to what previous systems achieved, which topped out at roughly two percent. While the study comes from Tavus itself and has not yet been independently verified, the gap between old and new is wide enough to suggest a meaningful shift in the technology's capability.
The underlying challenge in building convincing real-time video avatars has always been multi-layered. Latency, visual fidelity, and the ability to read and respond to conversational cues all have to work together without perceptible delay. Earlier systems tended to fall apart on at least one of these axes, producing responses that felt scripted or visually mismatched with natural human cadence. Griffin appears to address the integration of these signals more cohesively than what came before.
The practical implications span a range of industries - customer service, telehealth, sales, and companionship applications among them. However, the same capabilities that make Griffin useful also raise straightforward concerns about disclosure and consent. If nearly half of people in a controlled study cannot identify an AI interlocutor within a minute, the question of how and when users are informed they are speaking with a machine becomes more pressing. How platforms and regulators respond to that question will likely shape where this technology can be responsibly deployed.
