Tavus confirmed Griffin-Lite is not available to customers, citing deception capability as the explicit reason for the limited release while safety features are developed.
Following·Day 2·first covered 1 Oct 2026·7 sources
Tavus Griffin Convinces 48% of Live Participants It Is Human
The gist
A system that convinces roughly half of participants it is human in live video calls raises direct questions about identity deception in remote communication. Tavus itself acknowledges that the same properties enabling natural human-machine interaction also enable deception about AI identity, which is the stated reason it is withholding general customer access.
The full picture
Tavus launched Griffin, a real-time video AI that 48% of participants in a company-run 54-person study believed was a real human during a one-minute face-to-face video call, compared to 2.4% for the previous Phoenix-4.5 + Sparrow-2 + Raven-1 stack. Griffin scored 3.83 out of 5 on NVIDIA's VideoFDB benchmark for generation, below a human reference score of 3.92 but above the next-best published system at 2.80. Tavus calls Griffin the first 'Human Interaction Model,' a unified full-duplex system that processes video and audio simultaneously rather than chaining separate models for hearing, thinking, speaking, and animating. It reads facial expressions, tone, and pauses, can interrupt and be interrupted, and can change face, voice, gaze, and gestures mid-conversation. Griffin-Lite is currently a research preview for select testers only. Tavus has raised $70M in funding from investors including CRV and Sequoia Capital.
How it developed
Tavus launched Griffin, reporting a 48% human-pass rate in a 54-person live video call study and a 3.83 score on NVIDIA's VideoFDB benchmark.
Sources
Rohan Paul Twitter
The Neuron
ca.finance.yahoo.com
2 more sources
The daily email
Want this in your inbox?
I send one email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free