Tavus prepares a safety-focused early preview of Griffin for realistic AI interactions

A Tavus representative said the company is developing models for “pals”—AI employees, coworkers and companions. According to the representative, they are already being used to build AI assistants on Salesforce websites, continuing medical education tools, and AI tutors and healthcare workers. Tavus is targeting work that calls for human trust, emotional intelligence and visual interaction, where there are not enough people to meet demand.

Tavus describes Griffin as an early research preview available only to selected testers. The representative called it the company’s first product and its largest leap yet toward passing the Turing test. The company is still working on safety and on ensuring the model always discloses that it is AI, not human; a broader release was planned over the following couple of months. The representative acknowledged that more work is needed: in the study discussed, 15% of participants mistook a real person for AI.

According to the representative, Griffin can make small movements—such as shifting in a chair, gesturing and touching its face—but does not yet stand up and walk away, despite being technically capable of doing so. The company wants to make virtual employees and assistants more capable while making the interface less noticeable. The representative also described an “Oracle and Envoy” architecture: Griffin handles natural conversation and simulating human behavior, and can turn to external tools and models to do work in the background when a task goes beyond conversation.

The representative named piano instruction as one desired application. The model already has visual perception, they said, and users show it objects and play chess with it. But it cannot yet distinguish musical keys, making piano and guitar instruction difficult for now. The long-term ambition, according to the representative, is a personal tutor or trainer for everyone.