Tavus Griffin fools 48% of users in video calls

On 2 October 2026, Tavus unveiled Griffin, a video-based AI model designed for real-time interaction.

BigGo Finance reports that 48% of test subjects believed they were speaking to a real person during a video call. The model reacts to user inputs with lifelike latency and facial expression.

The release lands amid a busy week for generative video tools. Alibaba launched Wan3.0 on 24 August, following a $10 billion share sale. Black Forest Labs introduced FLUX 3 in July, a model capable of generating images and 20-second videos with audio. Griffin prioritizes interactive realism over static generation.

The Turing Test in Video

Griffin's performance in the video Turing test is the central metric of this release. Business Today reported that 48% of users could not tell the difference between Griffin and a human. This is a high success rate for a synthetic agent in a unmediated, real-time setting. Previous AI models often struggled with the nuances of live conversation, such as eye contact and speech timing. Griffin appears to have closed much of that gap.

"48% of test subjects mistook video call partner for a real human."

BigGo Finance

The implications for verification and trust are immediate. If nearly half of users cannot detect the AI, traditional cues for authenticity are failing. This raises questions for platforms that rely on video for identity verification, from banking to social media.

BeInCrypto noted that 26 of 54 people in their sample did not know they were talking to an AI. This aligns with BigGo Finance's 48% figure, suggesting a consistent pattern of deception. The models are becoming good enough to fool casual observers.

Context from Recent Releases

Griffin is part of a broader trend in AI video generation. On 10 August, LTX-2.5 was released, capable of generating 10-second videos from images in 6.8 seconds. It runs on Nvidia superchips and is open-weights. This speed and accessibility lower the barrier for creating realistic video content. Griffin adds a layer of interactivity that static generation models do not offer.

Alibaba's Wan3.0, launched in August, is another data point. The company raised $10 billion in a share sale before the release. This indicates significant capital investment in video generation capabilities. Tavus's Griffin distinguishes itself by focusing on the live video call format.

Black Forest Labs' FLUX 3, released in July, also pushes boundaries. It generates images and 20-second videos with audio. The inclusion of audio is a key feature, as sound adds another dimension to realism. Griffin takes this further by reacting to audio inputs in real-time. The combination of visual and auditory cues makes the deception more effective.

Verification and Ethics

The ease with which Griffin fools users raises ethical concerns. Misinformation and deepfakes become harder to detect when the AI is live and interactive. Platforms must develop new tools to verify the authenticity of video participants.

Regulators and tech companies are likely to respond to these developments. The ability to pass a video Turing test at a 48% rate is a milestone. It suggests that future models may reach even higher deception rates.

For now, Griffin acts as a benchmark for real-time AI video. Its performance highlights the rapid progress in synthetic media. As these tools become more common, the need for strong verification methods will only grow.