AI startup Tavus says its new model Griffin can fool people on video calls into thinking it is human. The company unveiled Griffin on October 1. It calls it the first Human Interaction Model. In Tavus’s own test, 26 of 54 people believed they were talking to a real person. That is 48 percent. Each test call lasted one minute. Tavus says its older system fooled only 1 of 41 people. The company also claims Griffin tops NVIDIA’s VideoFDB benchmark for AI video conversation. It scored 3.83 out of 5 on generation quality. Human recordings scored 3.92. Critics urge caution about the claims. The test was run by Tavus itself, not independent researchers. Griffin-Lite is available only to select testers for now. Tavus says it is holding back wider access until safety features are ready. The launch has sparked debate about AI deception risks. Regulators in the US and EU are already asking questions about disclosure rules for AI video callers.