Tavus Griffin AI fools nearly half of users in live video calls

A new artificial intelligence system designed for real-time face-to-face conversation deceived nearly half of the people who interacted with it into believing they were speaking to another human, according to data released by developer Tavus on October 1.

The model, called Griffin-Lite, convinced 26 out of 54 participants—48 percent—that they were talking to a real person during a one-minute video call. That represents a drastic leap from the company’s previous architecture, which fooled only one out of 41 testers. The older iteration relied on stitching together three separate models handling visuals, dialogue, and perception individually. Griffin consolidates those functions into a single system Tavus has branded a Human Interaction Model, engineered specifically to process facial expressions, conversational pauses, and spoken words simultaneously rather than sequentially.

Unlike traditional text-based large language models, Griffin operates in full-duplex mode. It listens, watches, and generates responses at the same time, mimicking the overlapping rhythm of a natural phone call rather than the turn-taking structure typical of current voice assistants. Tavus reports an average audio-to-video latency of 0.43 seconds when running on NVIDIA H100 chips, roughly half the delay of competing methods. In demonstrations, the system can visually track physical objects—such as guiding a user through solving a Rubik’s cube based solely on camera input—and pause naturally when the user stops speaking to think.

On NVIDIA’s VideoFDB benchmark, which evaluates live audiovisual conversation systems, Griffin-Lite scored 3.83 out of 5 on the generation track, measuring how natural and expressive the output appears. The next-highest scoring system reached 2.80, while the human reference baseline scored 3.92. On the perception track, which tests whether a model accurately interprets visual and auditory inputs, Griffin-Lite scored 3.73. That trails the human reference score of 4.20 but edges past the strongest non-human baseline of 3.44. Tavus states that NVIDIA conducted the evaluation independently.

The methodology behind the 48 percent deception rate differs significantly from established scientific protocols. Participants were recruited through what Tavus describes as an independent research platform and told they would be matched with another person to discuss their plans for the year. They were not warned that an AI might be involved. Only after the call concluded were they asked if they suspected their partner was synthetic. A community note appended to discussions of the findings on X flagged that the results lack independent verification and deviate from standard testing frameworks. Tavus noted that among the participants who did grow suspicious, most detected the artificial nature within the first 20 seconds.

This distinction matters because the original Turing test, proposed by British mathematician Alan Turing in 1950, requires judges to actively attempt to distinguish between a hidden human and a hidden machine. Text-only models have already demonstrated high deception rates under modified conditions. A University of California San Diego study found that OpenAI’s GPT-4.5 convinced judges it was human in 73 percent of interactions, though that experiment required prompting the AI to adopt a specific persona—an introverted, internet-savvy young person—and relied entirely on text. Griffin introduces synchronized facial animation and voice to the equation.

The capability to sustain realistic, interactive video deception carries immediate security implications. State-sponsored threat actors are already exploiting less sophisticated deepfake technology. In January, hackers linked to North Korea’s Lazarus Group, operating through the BlueNoroff subsidiary, used AI-generated video on Zoom and Microsoft Teams calls to impersonate trusted contacts. Victims were socially engineered into installing malware disguised as audio troubleshooting software. Security experts have warned since then that live video can no longer be treated as definitive proof of identity.

Corporate defenses currently rely on improvisation. In 2025, cryptocurrency exchange Kraken identified a suspected North Korean operative applying for a remote job after its security team deployed spontaneous, unscripted questions during an interview, such as requesting specific local restaurant names and government identification. The applicant faltered. As real-time generative models close the gap in latency and visual fidelity, these manual interrogation tactics will face diminishing returns.

Tavus, which closed a $40 million Series B funding round led by CRV in November 2025, is withholding Griffin-Lite from commercial release. The system remains restricted to select trusted testers who request access via a submission form on the company’s website. Tavus says it is collaborating with AI safety organizations to develop mandatory disclosure features before any public deployment.