This AI Avatar Fooled 48% of People in Job Interviews — Meet Tavus Griffin

Person on a video call (illustrative image)

Could you tell a real person from an AI on a video call? A Bay Area startup called Tavus says its new AI avatar model, Griffin, comes uncomfortably close. In a job-interview test, 48% of the 54 participants who chatted with the AI-generated interviewer believed they were talking to a real human — a result the company says shows its avatar can pass the Turing Test. Here is what Griffin does, what the test really proves, and why it matters.

What Is Tavus Griffin?

Griffin is a real-time AI avatar model that generates every pixel of every frame live from a single reference image, according to Tavus. That includes:

  • The face, arms and fingers
  • Natural breathing and blinking
  • Movement of the chair and the shadows it casts
  • The entire background scene

The avatars come in male and female looks. The big technical claim is that Griffin reacts in real time — it can interrupt a person, adapt to a new topic mid-conversation, and be interrupted without losing its train of thought.

AI avatar in a video meeting (illustrative image)

The Job Interview Test: 48% Fooled

According to Tavus, 54 participants each chatted with an AI-generated interviewer in a simulated job interview. The results:

  1. 48% believed the interviewer was human
  2. Those who thought it was human were 79% confident, on average, in that judgement
  3. Those who spotted the bot were 81% confident it was fake

The company presents this as a Turing Test milestone — evidence that a machine can now exhibit conversational behavior that is effectively indistinguishable from a human in this setting.

A critical caveat: these are the company's own claims from its own test. The sample size is small (54 people), the test was designed and reported by the company selling the technology, and no independent verification has been published. Treat it as an impressive demonstration, not a settled scientific result.

Why Small Details Matter

Tavus CEO and co-founder Hassaan Raza says the Griffin model fixes the subtle "reality-breaking" moments that used to give AI avatars away — like an avatar smiling while hearing bad news. His argument is that human communication is mostly non-verbal: expressions, tone, gestures and timing. Griffin's leap, he says, is getting those micro-behaviors right, not just the words.

That is a meaningful shift. Earlier avatar systems looked right but behaved wrongly — stiff, delayed, emotionally flat. If Griffin genuinely reacts in real time with appropriate emotion, the uncanny valley gets a lot narrower.

Video call with a virtual avatar (illustrative image)

The Obvious Risks

Technology that can convincingly fake a human on a live video call has serious implications:

  • Fraud: fake recruiters, fake executives on video calls, and more convincing social-engineering scams
  • Hiring: AI interviewers conducting real interviews — raising questions about fairness and disclosure
  • Misinformation: realistic fake people saying things on video that were never said by anyone

The same realism that makes Griffin impressive makes it a powerful tool for deception. As these avatars improve, verifying who is really on the other end of a call becomes a genuine security problem.

AI-generated digital avatar illustration (illustrative image)

What This Means for You

For most people, the near-term impact is simple: do not trust a face on a video call as proof of identity. Verify through a second channel — a known phone number, an official email, an in-person meeting — before acting on anything a "person" asks for on a call, especially money or sensitive information.

On the positive side, realistic avatars could make customer support, education and accessibility dramatically better — a patient, always-available human-like assistant in any language. The technology itself is neutral; the question is who controls it and how transparently it is used.

Disclaimer: Test results, technical claims and statistics in this article are Tavus's own claims as reported in October 2026; they have not been independently verified.



Post a Comment

Previous Post Next Post