Tavus Debuts Griffin, an AI Video Avatar That 48% of Testers Took for Human
Tavus says its Griffin model, which runs live video calls, convinced nearly half of study participants they were talking to a real person after one minute — up from a 2% ceiling on earlier systems.
Tavus has unveiled Griffin, a system it bills as the first Human Interaction Model, capable of holding a live video call while reading a caller's facial expressions, vocal tone and gestures. In a company-run study, 48 percent of participants said Griffin was a real human after a one-minute conversation. Tavus says prior conversational video systems topped out at 2 percent on the same measure, a jump the company is using to argue that real-time synthetic video has crossed a threshold users can no longer reliably detect.
The comparison figure matters as much as the headline number. Earlier avatars, including Tavus' own earlier generations, tended to break the illusion within seconds — a delayed mouth shape, a flat delivery, a gaze that drifted at the wrong moment. Griffin's pitch is that it closes those gaps in the same loop: it watches the person on the other end and adjusts its own expression, timing and intonation on the fly, rather than playing back a pre-rendered clip. Tavus has not published the study's methodology, sample size or the wording of the question put to participants, which leaves the 48 percent figure difficult to verify independently.
The company is positioning Griffin less as a content tool than as infrastructure for live conversation, a category that sits closer to telephony and customer support than to video editing. That framing puts Tavus in direct contention with a widening field of real-time avatar and voice vendors, several of which have shipped low-latency video agents in the past year. The differentiator Tavus claims is the interaction model itself — one network handling perception and generation together — rather than a pipeline of separate speech, vision and rendering components stitched together.
Why a one-minute call is the test that matters: most deployments Tavus is chasing are short. Support triage, sales qualification, appointment reminders and onboarding checks typically run under two minutes, and those are exactly the windows where a convincing avatar can replace a human without the conversation ever getting deep enough to expose its limits. Longer calls, where memory, consistency and unexpected questions pile up, remain the harder problem, and Tavus has not released figures for those.
The disclosure also lands in the middle of a broader fight over synthetic media. Regulators in the EU and several US states now require disclosure when consumers interact with AI systems, and a model that fools 48 percent of people in 60 seconds sharpens the case for watermarking or mandatory notice. Tavus has not said how Griffin labels itself on a call, or whether that is configurable by the customer deploying it.
What to watch: whether Tavus publishes the study in full, and whether outside researchers can reproduce the 48 percent figure under conditions the company does not control. The other open question is latency and cost at scale — a live avatar that reads expressions in real time needs far more compute per minute than a recorded clip, and the economics of running thousands of simultaneous calls will decide how quickly Griffin moves from demo to deployment.