Tavus Griffin AI Model Convinces 48% of Users They're Speaking to Real Humans on Video Calls

Tavus Griffin Redefines AI on Video Calls

AI startup Tavus introduced Griffin, a Human Interaction Model designed to conduct real-time video conversations with unprecedented realism1

1

2

2

. The AI model convinced 48% of participants across 54 calls that they were speaking with a real person, according to Tavus' internal research2

2

. This marks a dramatic leap from the company's previous system, which scored just 2.4% on the same video Turing test2

2

. Unlike traditional AI-generated video calls that rely on fragmented, turn-based systems, Griffin operates as a full-duplex platform, processing audio and video simultaneously while generating responses1

1

. The San Francisco-based company positions Griffin as the first of its kind, built specifically for face-to-face interaction rather than text-based exchanges3

3

.

How Human Interaction Model Achieves Hyper-Realistic Avatars

Griffin's architecture combines two engines working in tandem to create human-like interactions4

4

. One conversational model processes audio and video data to generate control signals for speech, emotions, facial expressions and gestures. A second generation engine translates those signals into voice and video output, producing 720p video in 320-millisecond increments from just a single reference image4

4

. The system achieves an average audio-to-video delay of 0.43 seconds on NVIDIA H100 chips, which Tavus claims is half that of the next fastest method2

2

. During demonstrations, the Griffin avatar displayed remarkably natural behaviors including swinging in a chair, reacting to conversational interruptions, incorporating "ums" and "ahs" into responses, and even waving when someone entered the frame1

1

. The AI model can clone voices using approximately 10 seconds of audio data4

4

.

NVIDIA VideoFDB Benchmark Results Validate Performance

On NVIDIA VideoFDB benchmark, which tests live audio and video conversation capabilities, Griffin-Lite scored 3.83 out of 5 on the generation track that grades naturalness and expressiveness of responses2

2

. The human reference scored 3.92, while the next-best system achieved only 2.802

2

. NVIDIA ran this evaluation independently, according to Tavus2

2

. On the perception track, which measures how well a model understands what it sees and hears, Griffin-Lite scored 3.73 compared to the human reference of 4.20 and the strongest baseline of 3.442

2

. This represents the highest score among 15 models tested4

4

. The company demonstrated Griffin coaching someone through a Rubik's cube based on visual input, waiting naturally when the person paused to think2

2

.

Ethical Concerns and Digital Trust Issues Emerge

The unveiling of Griffin sparked immediate backlash over ethical concerns and potential misuse. Former Congressman Tom Malinowski stated, "It should be illegal -- plain and simple -- for any AI to impersonate a human being without a clear and visible disclaimer that it is, in fact, AI"1

1

. Bahrad Sokhansanj, a senior research fellow at the Institute for Law & AI, called it "an atrocity" that "exemplifies the reckless culture that has become all too prevalent in tech"1

1

. Paul Novosad, Professor of Economics at Dartmouth, added that "working on camouflaging AI use should be viewed as unethical"1

1

. Concerns center on potential use in scams and fraud, particularly given that bad actors have already used AI-generated voices in telephone scams1

1

. North Korea-linked hackers previously used deepfakes on Zoom or Teams calls to pose as trusted contacts, tricking victims into installing malware2

2

.

Tavus CEO Hassan Raza Addresses AI Safety Measures

Tavus CEO Hassan Raza responded to criticism on X, stating, "We're working with partners on safeguards and systems for disclosure, as well as inviting discussions with officials around wider regulation and safe use"1

1

. He emphasized that "people will always know they're interacting with AI, while providing an interface that removes the need to 'speak computer'"1

1

. The company is limiting access to Griffin-Lite to select trusted testers while developing disclosure features and working with AI safety organizations2

2

5

5

. Tavus claims Griffin is intended for "low access or high cost" work in industries like education or healthcare, such as tutoring services where AI adapts to student learning styles or health experts providing medication guidance with greater empathy than text-based ChatGPT1

1

. The company raised a $40 million Series B in November 2025, led by CRV2

2

.

Research Methodology Faces Scrutiny

The 48% success rate comes from Tavus' own research page, where participants were told they would be matched with another person for a one-minute video call about what they were looking forward to this year2

2

. Only after the call ended were participants asked whether it had crossed their mind that their partner might not be real2

2

. This differs from the classic Turing test proposed in 1950, where a judge knowingly evaluates both a human and machine to determine which is which2

2

. A community note on X flagged that results are not independently verified and do not follow a standard protocol2

2

. The one-minute duration and small sample size of 54 people raise questions about whether longer conversations would yield similar results1

1

4

4

. Tavus reports that participants who grew suspicious typically did so within 20 seconds2

2

.

Customer Support and Job Interview Applications Raise Questions

Tavus points to use cases including interview practice, tutoring, and customer support where natural conversation matters3

3

5

5

. The model's ability to process speech and visual signals could make it useful for interactive training and simulations3

3

. However, using AI to impersonate a candidate in actual job interviews raises serious questions around disclosure, consent and assessment integrity3

3

. Some employers already prohibit candidates from using generative AI during live interviews3

3

. Security researchers from Kraken flagged a suspected North Korean job applicant after asking spontaneous questions about government ID and local restaurants, which the candidate struggled to answer2

2

. Griffin-Lite remains unavailable to general customers, with eligible testers able to request access by submitting a form on the Tavus site2

2

5

5

.