Artificial intelligence startup Tavus has developed a video AI model named Griffin that successfully convinced 26 out of 54 participants (48%) on live video calls that they were conversing with a real human.
Key Takeaways
- High Realism: Griffin passed the video Turing test by fooling 48% (26 of 54) of live participants, up from 1 of 41 (<3%) in previous system tests.
- Benchmark Leader: Scored 3.83 out of 5 on NVIDIA's VideoFDB benchmark, approaching the human reference score of 3.92 and surpassing a Gemini 2.5 and Anam combo score of 2.80.
- Fast Voice Cloning: Griffin-Lite generates conversational speech and can clone a speaker's voice using approximately 10 seconds of audio.
- Controlled Access: Tavus is withholding the model from general commercial release due to deception risks, opening access only to select trusted testers.
Breakthrough in Video Turing Test Performance
Tavus co-founder and CEO Hassaan Raza explained that human communication relies heavily on expression, tone, and timing rather than text prompts alone. Unlike traditional cascade systems that alternate between speech recognition, language models, and voice synthesis, Griffin reads video, verbal inputs, and nonverbal cues together every fraction of a second. This architecture allows the model to process interruptions smoothly and offer natural feedback like nodding or saying "mm-hm" while a user is actively speaking.
In testing, participants anticipated a one-minute live video conversation with another person about their year ahead. More than half reported that the possibility of interacting with an AI never crossed their minds, though suspicious users identified the model within the first 20 seconds. The system, designated as a Human Interaction Model (HIM), also clones a target voice using roughly 10 seconds of sample audio.
Fraud Risks and Restricted Deployment
Because the technology presents severe deception risks, Tavus is withholding Griffin from general public access. The preview version, Griffin-Lite, is currently restricted to vetted testers while Tavus collaborates with AI safety organizations and builds built-in disclosure mechanisms.
These safety precautions come as AI agents and automated bots drive significant online activity. Blockchain security firm TRM Labs reported that criminal AI adoption expanded 40% year-on-year in 2026, driven primarily by fraudulent schemes. As security experts issue warnings regarding rising AI cyber attacks, hyper-realistic video models could undermine conventional identity verification practices.
Why It Matters
The emergence of real-time video models that pass human perception tests creates urgent challenges for digital identity and financial compliance. If video and voice authentication can be spoofed using 10 seconds of audio, traditional video Know-Your-Customer (KYC) onboarding used by crypto exchanges will become increasingly vulnerable. Moving forward, crypto platforms and financial institutions will be forced to rely more heavily on cryptographic proofs, hardware security keys, and on-chain identity protocols rather than visual human confirmation.
