Tavus Unveils Griffin Dialog Model After 48% of Test Callers Took It for Human
Tavus is holding back general availability of a model it says people could not reliably tell from a human, which puts the disclosure features, not the dialogue quality, at the center of when Griffin actually ships.
Reporting from 1 source: ASCII.jp.
US company Tavus announced Griffin on October 1, a dialogue model that reads audio and video at the same time. In an internal test, 26 of 54 participants recruited in the US and Europe who held a one-minute video call with the research preview Griffin-Lite judged the other party to be a real human. Tavus is working on safety measures, including features that indicate it is AI, before general availability.
The test ran on Griffin-Lite, the research preview, not the announced Griffin model itself. Participants were told the other party was another participant, and 26 of 54 judged it to be human. Griffin is built to decide whether to speak, wait, or backchannel at intervals under one second while it listens to and watches the conversation, and it generates facial expressions, gaze, and gestures in step with the exchange. Only early testers can use Griffin-Lite now. Tavus says human-indistinguishable dialogue models require cautious release and will reach general availability after its safety features are in place.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.