Tavus Griffin AI Model Fools 48% of People Into Thinking They're Talking to a Real Human

Tavus Griffin Convinces Nearly Half of Test Participants They're Speaking With Humans

San Francisco-based AI startup Tavus has unveiled Griffin, what it calls the first Human Interaction Model, and the results are striking. In company-conducted research, 48% of 54 participants believed they were speaking with a real person during one-minute video calls with the AI model1

1

. This marks a dramatic leap from Tavus' previous system, which convinced only 1 out of 41 people, scoring just 2.4% on the same test1

1

.

Participants were told they would be matched with another person for a conversation about what they were looking forward to this year. Only after the call ended were they asked whether it had crossed their mind that their partner might not be real1

1

. Those who grew suspicious typically did so within 20 seconds1

1

. The results come from Tavus' own research page using an independent research platform, and a community note on X has flagged that the findings are not independently verified and do not follow a standard protocol1

1

.

Real-Time Processing Eliminates the Awkward AI Pause

Unlike conventional chatbots that operate through a relay process, Griffin is built for human-like face-to-face conversations2

2

. The AI model can listen to spoken questions, interpret visual signals, and respond with speech, facial expressions, gestures, and body movements2

2

. What sets Griffin apart is its ability to eliminate the characteristic pause that defines most AI-driven human interaction systems3

3

.

Griffin operates as a video-to-video system that listens and observes as it speaks3

3

. Subsecond by subsecond, it determines whether to speak, nod, use acknowledgments like "mm-hm," or remain silent3

3

. This full-duplex capability means it listens, watches, and talks simultaneously, like a phone call rather than a walkie-talkie1

1

. The system can handle interruptions, shifts in tone, and visual cues shown on screen, allowing it to react while the conversation is still happening2

2

.

Griffin-Lite Tops NVIDIA VideoFDB Benchmark With Subsecond Delays

On NVIDIA's VideoFDB benchmark, a test of live audio and video conversation, Griffin-Lite ranks first1

1

. The generation track, which grades how natural and expressive an AI model's responses are, gave Griffin-Lite a score of 3.83 out of 51

1

2

2

. The next-best system scored 2.80, while the human reference scored 3.921

1

2

2

. Tavus says NVIDIA ran the evaluation independently1

1

.

The perception track, which measures whether an AI model understands what it sees and hears, shows Griffin-Lite scored 3.73 against 3.44 for the strongest baseline, while the human reference hit 4.201

1

. Audio-to-video delay averages 0.43 seconds on NVIDIA H100 chips, the kind used in AI data centers, which Tavus claims is half that of the next fastest method1

1

3

3

. In a demonstration video, Griffin coaches someone through a Rubik's cube based on what it sees in their hands and waits when the person goes quiet to think1

1

.

Two-Engine Architecture Powers Natural Interactions

Griffin's architecture relies on two engines operating simultaneously3

3

. One conversational model consumes audio and video data and produces control signals for speech, emotions, facial expressions, and gestures. Another generation engine translates those control signals into voice and face animations3

3

. The voice cloning system can replicate voices based on approximately 10 seconds of audio data, while the video generation creates 720p video in 320-millisecond increments based on just one picture reference3

3

.

This represents a significant departure from Tavus' previous system, which stitched together three separate models—one each for visuals, dialogue, and perception1

1

. The integrated approach allows Griffin to interpret timing, intonation, and camera footage without the information loss that occurs when multiple systems hand off data to each other3

3

.

Digital Trust Concerns Mount as Deepfakes Proliferate

The technology arrives at a moment when scammers already exploit video calls for malicious purposes. In January, North Korea-linked hackers used deepfakes on Zoom or Teams calls to pose as trusted contacts1

1

. Security researchers attribute the intrusion to BlueNoroff, a Lazarus Group subsidiary, with victims talked into installing malware disguised as an audio fix1

1

. David Liberman, co-creator of Gonka, a decentralized network for AI computing, stated in that report that photos and video can no longer be trusted as proof that something is real1

1

.

Companies have begun improvising defenses. In 2025, Kraken flagged a suspected North Korean job applicant after its security team asked spontaneous questions, like requesting government ID and the names of local restaurants1

1

. The candidate struggled to respond1

1

. The ability of Griffin to handle interruptions and respond to visual cues could make such verification methods less effective, though the implications for job interviews raise questions around disclosure, consent, and assessment integrity2

2

. Some employers already prohibit candidates from using generative AI during live interviews2

2

.

Griffin-Lite Remains Restricted as Tavus Develops Safety Measures

Griffin-Lite is not available to customers and is limited to select trusted testers as a research preview1

1

4

4

. Tavus says it is working on disclosure features and with AI safety organizations before a public release1

1

4

4

. The company acknowledges that the same characteristics making Griffin natural are precisely what make users think it's not an AI model3

3

.

Trusted testers can request access to Griffin-Lite by submitting a form on the Tavus site1

1

. Currently, Tavus customers build upon the company's Phoenix, Raven, and Sparrow products3

3

. Tavus raised a $40 million Series B in November 2025, led by CRV1

1

.

Applications Span Customer Support, Tutoring, and Practice Interviews

Tavus points to potential use cases including practice interviews, tutoring, and customer support2

2

4

4

. The AI model's ability to process speech and visual signals could make it useful for interactive training and simulations where natural conversation matters2

2

. Users can interact with Griffin naturally through conversation rather than learning specific commands or prompts4

4

.

The development signals a shift toward AI systems that don't simply answer questions but participate in more human-like, two-way video interactions2

2

. However, the narrow parameters of the 48% finding—one-minute conversations with 54 people in a company-conducted study—leave questions about how Griffin would perform over longer interactions or in more challenging scenarios3

3

. What happens when conversations extend to an hour remains unknown, and whether this represents a genuine Turing Test pass or simply a well-executed demonstration continues to be debated3

3

.