Tavus says its Griffin video model passed for human 48% of the time
Tavus says 48% of study participants took its Griffin model for a real person after a one-minute video call, against 2% for earlier AI systems.
Tavus, a San Francisco company founded in 2020, has introduced Griffin, which it calls its first "Human Interaction Model". The system holds real-time video calls and processes speech, facial expressions, tone of voice, gestures and pauses. The Decoder reported the launch on Thursday, citing Tavus's own research report.
In a Tavus study, 48 percent of participants believed Griffin was a real person after a one-minute call, according to The Decoder. The company says earlier AI systems reached 2 percent on the same measure. TestingCatalog, citing Tavus, put the figure at 45 percent of 120 participants, so the exact number differs between the two accounts.
https://x.com/mark_k/status/2105718951778955317
Tavus also cites an Nvidia benchmark. On Nvidia's Video Full-Duplex Benchmark, TestingCatalog reports, Griffin Lite scored 3.83 out of 5 against 3.92 for real humans. The Decoder says the best earlier AI scored 2.80. Tavus describes the Nvidia test as independent. We have not seen Nvidia's own results.
Griffin is not generally available. Tavus is offering Griffin Lite to select testers as a research preview, The Decoder reports. A more capable version will wait until what the company calls safety concerns are addressed. Tavus suggests uses such as tutoring, practising difficult conversations and camera-based tech support.
TestingCatalog says Griffin learns conversational behaviour from people and can interrupt mid-sentence or respond when someone enters the frame. These are Tavus's claims, drawn from its own study. Nobody outside the company has said they tested the model, and the Turing-test framing rests on a one-minute call.