Artificial intelligence startup Tavus unveiled a real-time conversational video model called Griffin on October 1, 2026. The model convinced 48 percent of study participants that they were interacting with a real human during live video calls. This milestone marks a dramatic leap in generative video technology, crossing a threshold that researchers have long viewed as the video Turing test for synchronous AI communication.
The controlled test involved 54 participants engaging in one-minute video calls with the Griffin-Lite variant. Out of that cohort, 26 users reported believing their conversation partner was a real human being. This represents a massive increase from the 2.4 percent success rate achieved by Tavus’s previous Phoenix-4.5 model across 41 participants, highlighting a nearly twenty-fold improvement in perceived realism within a single development cycle.
How Griffin Operates as a Full-Duplex System
Legacy avatar systems chain separate pipelines for speech recognition, language processing, speech synthesis, and facial animation. This results in awkward pauses and monotonous responses. In contrast, Griffin operates as a unified, full-duplex system that simultaneously processes visual input, listens to audio, and speaks in real time.
Operating on NVIDIA H100 graphics processing units, Griffin reduces audio-to-video latency to an average of 0.43 seconds. This low-latency architecture allows the generated avatar to:
- Interrupt users mid-sentence without losing conversational context.
- Alter gestures, head tilts, and facial expressions organically.
- Inject verbal backchannels such as “mm-hm” while the human user is still talking.
- Evaluate temporal understanding, recognizing when a pause signifies thought rather than the end of a turn.
Independent Benchmarking on NVIDIA VideoFDB
While Tavus’s internal user study provided the headline 48 percent deception rate, external validation came from NVIDIA’s Video Full-Duplex Benchmark (VideoFDB). This independent industry benchmark evaluates how naturally a model produces conversational behavior and how accurately it perceives human cues.
On NVIDIA’s evaluation tracks, Griffin-Lite achieved exceptional scores:
- Generation Track: Scored 3.83 out of 5, placing it just 0.09 points behind the human reference score of 3.92 and outperforming competing commercial systems by over a full point.
- Perception Track: Scored 3.73 out of 5, making it the most perceptive real-time model on the benchmark among 15 evaluated systems.
- Takeover-Rate Alignment: Reached 62.8% on generation and 73.8% on perception, tightly matching human conversational timing.
Safety Concerns, Enterprise Risks, and Release Plans
Despite the technological breakthrough, Tavus is withholding the full Griffin model from general commercial release. The company has restricted access to a research preview known as Griffin-Lite for select trusted testers while it develops robust safety and disclosure safeguards.
In its official announcement, Tavus acknowledged the inherent risks of its creation. The company stated that the exact properties making Human Interaction Models powerful communication tools also allow them to convincingly deceive humans. Cybersecurity experts and industry observers have raised urgent warnings regarding potential malicious applications. These range from sophisticated video phishing and executive impersonation to fraudulent financial authorizations.
Regulatory frameworks like the European Union’s AI Act mandate strict transparency and watermarking for synthetic media. Because of this, the development of built-in cryptographic verification and disclosure features will dictate how safely these advanced video avatars transition into enterprise deployment.
Frequently Asked Questions
What is the Tavus Griffin AI model?
Griffin is the first Human Interaction Model (HIM) developed by Tavus. It is a full-duplex, video-to-video conversational system that processes visual and audio inputs simultaneously to generate realistic, real-time video responses from a single reference image.
How many people were fooled by Griffin in testing?
In a blind study of 54 participants who engaged in one-minute video calls, 48 percent (26 users) believed they were talking to a real human rather than an AI avatar.
What is the Video Turing test achieved by Griffin?
The video Turing test measures an AI model’s ability to maintain a synchronous, face-to-face video conversation without the human participant realizing they are interacting with a machine.
How does Griffin differ from older AI avatar technologies?
Older avatar systems relied on cascaded pipelines-transcribing speech, generating text, synthesizing audio, and animating a face sequentially-which caused noticeable lag and dead air. Griffin processes perception, decision-making, and generation concurrently at sub-second intervals.
When will Griffin be available to the public?
Tavus has not released a definitive public launch date. The company is currently offering a limited research preview called Griffin-Lite to trusted testers while developing mandatory safety, watermarking, and disclosure mechanisms to prevent misuse.
