Tavus's New AI Avatar Griffin Fooled 48% of People Into Thinking It Was Human

The Turing Test just got an avatar

Could you tell the difference between a human interviewer and an AI one? Bay Area startup Tavus is betting you can't. Its new AI model, Griffin, powered an AI interviewer that fooled nearly half the people it spoke with — a result the company says means it clears the Turing Test.

48% couldn't spot the bot

Tavus ran 54 participants through a job-interview conversation with the Griffin-powered avatar. 48% of them believed they were talking to a real human. The striking part: those who thought the interviewer was human were 79% confident about it — and those who thought it was a bot were 81% confident. Both sides felt sure, and half the room was wrong.

Every pixel is generated live

Unlike older AI avatars that play back pre-rendered clips, Griffin generates every pixel of every frame in real time from a single reference image. The face, arms, and fingers — but also the chair's movement, the shadows it casts, and the background. The avatars come in male and female looks, and they breathe and blink naturally.

The realism upgrades go beyond pixels. Griffin no longer smiles while hearing bad news, it reacts to real-time data, and it can interrupt a person — or be interrupted — without losing its train of thought. In demos, the avatar taught a participant to solve a Rubik's Cube and played a round of Simon Says, responding to visual cues as they happened. CEO and co-founder Hassaan Raza said humans are wired to communicate face to face — through expressions, tone, gestures, and timing — and Griffin is designed to speak that language.

Why it matters

The obvious uses are hiring and education: Tavus says the models could serve students in schools or as tutors. The flip side is harder to ignore. If an AI avatar can sail through a job interview undetected, the same technology can impersonate anyone on a video call. The uncanny valley isn't just closing — in short, scripted interactions, it may already be gone.

The take

Griffin's 48% isn't proof the Turing Test is "solved" — it was the company's own test, on 54 people, in one setting. But the direction is unmistakable: real-time, interactive, interruptible AI faces are here, and every pixel is generated on the fly. The question isn't whether AI avatars can fool us anymore. It's what happens when they do it routinely.