Observed facts: Tavus says Griffin-Lite is a research preview available to a select group of early testers. Tavus reports that 48% of participants in a live one-minute video-call study thought they had spoken with a real person, compared with a maximum 2% pass rate for its earlier systems.
Mechanism: Tavus describes Griffin as a full-duplex video-to-video system: perception, deciding whether and how to respond, and expressive speech and video generation operate at the same time. The proposed mechanism is continuous multimodal interaction rather than a sequence of transcription, text response, and synthesized output.
Why now: Griffin-Lite is available to early testers now, and Tavus says a wider release of a more powerful model will follow. The preview makes the interaction design and its evaluation a live question rather than a purely prospective research claim.
Implications: If the design works beyond demonstrations, conversational timing and nonverbal cues may become central product capabilities for real-time AI interfaces. The reported result also makes disclosure and user expectations important: perceived human-likeness is not itself evidence of general capability or safe use.
Evidence boundary: The supplied public source is Tavus’s own product and research announcement. Its study outcome, comparison with earlier systems, and performance framing are company-reported; the supplied material does not provide an independent replication.
Watch condition: Watch for broader access and independent evaluation of the one-minute call result under a disclosed study protocol.
The next useful evidence is whether the preview’s reported behavior survives broader testing and whether user-facing safeguards match the system’s human-like presentation.