The line between human and machine has just become significantly more blurred. AI research lab Tavus announced that in a recent study, nearly half of all participants could not tell that its new AI video avatar, Griffin, was not a real person.
What is Tavus and Griffin AI?
Tavus is
a US-based AI research company focused on creating what it calls the 'human layer of AI'. Their mission is to make interactions with artificial intelligence feel as natural as speaking with a person face-to-face. Their latest creation, Griffin, is a sophisticated AI model designed for real-time video conversations. Unlike previous systems that often feel clunky, Griffin is a 'Human Interaction Model'. It processes speech, visual cues like facial expressions, and non-verbal signals like pauses all at once, allowing it to see, hear, and respond in a unified, fluid manner. This simultaneous processing allows the AI avatar to do things like interrupt, react to being interrupted, and change its expression mid-sentence, just as a person would.
The Study That Made Headlines
The headline-grabbing claim comes from a study conducted by Tavus itself. The company recruited 54 participants for what they were told would be a one-minute video call with another participant to discuss their plans for the year. In reality, they were speaking with an avatar powered by a research preview of the AI, called Griffin-Lite. After the brief conversation, they were surveyed, and 26 of the 54 participants—or 48%—stated they believed their conversation partner was a real person. This result marked a massive leap from Tavus’s previous AI model, which fooled only 2.4% of participants in a similar test. While the company hailed this as the first model to pass a real-time 'video Turing test', it's important to note the study's context: it was a small sample size, the calls were very short, and the study was not conducted by an independent third party.
A New Benchmark in AI Interaction
Beyond its own study, Tavus’s Griffin-Lite also performed impressively on a more standardized test: NVIDIA's Video Full-Duplex Benchmark, which evaluates live audio-visual conversational AI. On this benchmark, Griffin-Lite ranked first among evaluated systems for both its ability to perceive human cues and the natural quality of its own generated responses. Its generation score was 3.83 out of 5, remarkably close to the human reference score of 3.92 and significantly ahead of the next-best AI system, which scored 2.80. This suggests that even under more formal evaluation, Griffin represents a major advance in creating AI that not only understands conversation but can participate in it with a new level of fluency and realism.
The Promise: A New Business Interface
The potential applications for such human-like AI are vast. Companies are already envisioning using these 'PALs' (agentic AI humans) for everything from customer service and sales to education and recruitment. Imagine an AI sales representative that can engage tens of thousands of website visitors a day with natural, face-to-face conversations, or an AI tutor that can patiently explain a concept with human-like expressions and gestures. For businesses, this technology promises to deliver personalized, scalable human touchpoints that were previously impossible. Salesforce, for example, has already used Tavus technology to build an AI sales rep that it reports has increased meeting bookings six-fold.
The Peril: Deception and Disinformation
However, the same technology that offers such promise also opens the door to significant ethical risks. The fact that nearly half of testers were deceived has raised immediate concerns about the potential for misuse, including sophisticated scams, fraud, and the spread of disinformation. Critics argue that AI this convincing should be illegal without clear, unmissable disclaimers. The technology could be used to impersonate individuals, create fake testimonials, or manipulate people in deeply personal ways. Recognizing this danger, Tavus has stated that Griffin is too powerful for a public release at this stage and is working on safety features and disclosure protocols before making it widely available.
