An American company named Tavus has introduced an AI model called Griffin, which it claims is the first to pass what it calls the "video Turing test." This concept was first proposed by mathematician Alan Turing in 1950 as a way to assess whether a machine can mimic human conversation so closely that it becomes impossible to tell the difference. Tavus says Griffin can be used for video calls, generating in real time not only its responses but also its appearance, voice, and expressions. The AI listens and observes the user, reacting to gestures, expressions, or interruptions as the conversation unfolds. Tavus describes Griffin's method as a "video-to-video" approach in "duplex" mode, meaning that the exchange of information happens simultaneously in both directions. This allows the AI to perform actions such as nodding during the user's sentence or adjusting its response if the user interrupts. Starting from a reference image, Griffin creates the entire scene, including body movements, shadows, and background settings, in real time. In one demonstration, Griffin joked about its own "breathing" to ease the load on the user's device and teased another participant about his sweater. Tavus tested Griffin-Lite, an early version of the model, with 54 people who had one-minute conversations. About 48 percent of them thought they were speaking with a human, compared to only about 2.5 percent with the company's previous system. However, the test had limitations, such as a small sample size and short interactions, and the participants were not explicitly trying to detect if they were talking to an AI. Tavus acknowledges the potential for deception and says it aims to make interactions more natural without making the AI appear human. The company is working on ways to clearly indicate that Griffin is artificial before making it available to customers. For now, Griffin-Lite is only accessible to a few testers in a research phase. The company has not announced a timeline or cost for a public release, but it promises a quick launch once security concerns are addressed. In addition, Griffin-Lite performed the best among models tested in VideoFDB, a benchmark developed by Nvidia to evaluate the naturalness of AI interactions. This includes checking if the AI responds appropriately to silence, eye contact, or changes in expression. However, the scores are assigned by another AI based on a predefined evaluation system. While Griffin-Lite scored higher than other AI models, it still lagged behind human interactions in the comparison. Some competing models also responded more quickly.