Tavus Griffin-Lite AI video fools 48% of callers into thinking it's human

Tavus's Griffin-Lite model convinced 48% of test participants they were talking to a real person, up from 2.4% for the previous system. On NVIDIA's VideoFDB benchmark it scored 3.83 out of 5 versus 3.92 for humans. World responded by pointing to World ID for Zoom as a proof-of-human safeguard.

• 48% mistook the AI for a human, up from 2.4% for the prior Tavus model
• Average confidence among fooled callers was 79%; 26 of 54 said the partner was real
• VideoFDB: 3.83 of 5 vs 3.92 for humans, closing 92% of the gap
• 720p video in 320 ms chunks, 0.43 s audio-to-video latency on H100