Tavus is an AI company building systems that can see, hear, understand and respond in real time. Its new Griffin model crosses an important threshold: it can hold a live video conversation with human-like timing, expressions and reactions. In a Tavus study, 26 of 54 people — 48% — thought Griffin-Lite was a real person, compared with 2.4% for the earlier system.
Griffin does more than display a talking face. It listens while speaking, notices pauses and interruptions, reads voice and facial cues, reacts to what appears on screen and changes its response in real time. It scored 3.83 out of 5 on NVIDIA’s VideoFDB benchmark, close to the human reference score of 3.92 and ahead of the next published AI system at 2.80.
But Griffin has not become human, nor has it cleared every human test. The result came from a small, company-run one-minute study, and it measures whether people were fooled—not whether the model truly understands, feels or reasons like a person. Its promise is clear in tutors, customer support, sales training, healthcare education, accessibility and companions for isolated or elderly people; its dangers include impersonation, scams, manipulation and deepfake relationships.
This is likely the beginning of a new AI race: machines that do not merely answer questions, but participate in conversations. Tavus is keeping Griffin-Lite with trusted testers while it develops disclosure and safety measures before a wider release. The future may bring helpful digital people—but every video call will need proof of who, or what, is actually on the other side.
WHEN AI CAN LOOK HUMAN, TRUST MUST BECOME VERIFIABLE.
Sanjay Sahay
Have a nice evening.