Tavus AI avatar fools 48 percent of test subjects on one-minute video call
Tavus has unveiled Griffin, a real-time AI video avatar that processes facial expressions, tone of voice, and gestures. In company tests, 48 percent of participants believed it was a real person after a one-minute call, up from a previous high of two percent.
Tavus has introduced Griffin, an AI system the company describes as the first «Human Interaction Model», capable of conducting real-time video calls while processing a participant's facial expressions, tone of voice, and gestures. In a study conducted by Tavus, 48 percent of participants believed Griffin was a real person after a one-minute call, a figure that far exceeds the two percent achieved by previous systems.
The result marks a significant shift in how convincingly artificial agents can pass as human in live, interactive settings. Earlier generations of conversational AI typically revealed their synthetic nature within seconds, through unnatural pauses, flat vocal delivery, or a lack of responsive body language. Griffin is designed to overcome those limitations by reading and reacting to visual and auditory cues in real time, allowing it to sustain the rhythm of a natural conversation.
Tavus has not disclosed the full methodology of the study, including the number of participants, how they were recruited, or the specific questions they were asked after the call. The company also has not released independent verification of the 48 percent figure. Even so, the reported jump from two percent to nearly half of all test subjects suggests that the technology has crossed a threshold that raises fresh questions about trust, consent, and the authenticity of digital communication.
The development arrives as video calls have become a default channel for job interviews, remote work, customer service, and even medical consultations. A system that can convincingly impersonate a human participant in that environment could be used for legitimate purposes, such as scaling personalized support or training, but it could also be deployed for deception. The same capabilities that make Griffin effective at maintaining eye contact and mirroring a caller's tone could be used to create fraudulent job candidates, fake customer representatives, or fabricated witnesses in online proceedings.
For religious and cultural institutions that increasingly rely on digital platforms for pastoral care, education, and community building, the implications are equally complex. A convincing AI avatar could extend the reach of a small staff or offer round-the-clock companionship to isolated individuals. It could also blur the line between genuine human presence and synthetic interaction, particularly in settings where empathy and moral accountability are expected. The question of whether a machine can truly listen, forgive, or accompany someone in grief is not merely technical but theological and ethical.
Tavus is positioning Griffin as a tool for businesses and developers rather than a consumer product, though the company has not detailed specific launch partners or pricing. The absence of independent testing means the 48 percent figure should be treated as a company-reported result rather than a peer-reviewed finding. Still, the direction of travel is clear: AI systems are becoming harder to distinguish from humans in short, real-time exchanges, and the burden of verification is shifting to the people on the other side of the screen.
Regulators in the European Union and elsewhere are already examining disclosure requirements for AI-generated content, but most rules focus on text, images, and pre-recorded video. Live, interactive avatars occupy a gray area. As the technology improves, the practical question for users may be less about whether they can spot an AI and more about whether they have any reliable way to know who, or what, they are talking to.
12
