AI startup Tavus announced on October 1 that its new model, Griffin, convinced 48% of people who engaged in live video calls that they were interacting with a human. The company describes Griffin as the first Human Interaction Model, designed to understand and generate face-to-face conversation by processing expressions, pauses, and words simultaneously. In testing, the Griffin-Lite version faced 54 participants, 26 of whom believed their partner was real after a one-minute call about future plans. This represents a significant jump from Tavus' previous system, which achieved only a 2.4% success rate among 41 participants.
The results, published on Tavus' own research page, indicate that suspicious participants typically realized the deception within 20 seconds. On Nvidia's VideoFDB benchmark, Griffin-Lite ranked first with a generation score of 3.83 out of 5, compared to 2.80 for the next-best system and 3.92 for a human reference. The model operates full-duplex, listening, watching, and talking at once, with an average audio-to-video delay of 0.43 seconds on Nvidia H100 chips. Currently, Griffin-Lite is limited to select trusted testers while Tavus develops safety measures and disclosure features. The company raised $40 million in a Series B led by CRV in November 2025.
The reported ability of Tavus' Griffin model to pass as human in unstructured video interactions marks a critical escalation in the realism of synthetic media. Unlike text-based benchmarks where models like GPT-4.5 have previously demonstrated high deception rates, this development integrates visual and auditory cues in real-time, lowering the barrier for sophisticated social engineering attacks. The rapid detection time of approximately 20 seconds suggests that while the initial impression is convincing, sustained interaction may still reveal artifacts, yet this window is sufficient for many fraud scenarios involving quick verification or malware installation prompts.
From a security and regulatory standpoint, the gap between current defensive capabilities and emerging generative tools poses immediate risks to institutional trust. As deepfake technology becomes more accessible and realistic, traditional identity verification methods relying on video calls are increasingly vulnerable. Organizations must anticipate a shift toward multi-factor authentication and behavioral analysis rather than visual confirmation alone. Furthermore, the lack of independent verification for these specific performance metrics highlights the need for standardized, third-party auditing of AI safety claims before widespread deployment.
