Tavus says its Griffin model fooled 48% of video call partners into thinking it was human
Decrypt reports Tavus says 26 of 54 people who had a one-minute video call with its Griffin-Lite model believed the partner was a real person.
Why it matters
The results come from Tavus itself and have not been independently verified, so the 48% figure carries the weight of a company demonstration rather than an audited measurement. It touches on how regulators and platforms define identity verification and impersonation in voice and video, and the company restricts Griffin-Lite to select trusted testers pending safety work, which limits near-term availability.
Key facts
26 of 54 participants, or 48%, believed Griffin-Lite was a real person after a one-minute video call.
48 %Reported by Decrypt
Tavus' previous system drew a single believer out of 41 participants, a 2.4% rate on the same test.
2.4 %Reported by Decrypt
Griffin-Lite scored 3.83 out of 5 on generation on Nvidia's VideoFDB benchmark, against 2.80 for the next-best system and 3.92 for a human reference.
3.83Reported by Decrypt
Tavus calls Griffin its first Human Interaction Model, designed to read expressions and pauses as well as words.
Reported by Decrypt
Participants were told they would be matched with another person for a one-minute call and were asked at the end whether the partner might not be real.
Reported by Decrypt
Griffin-Lite is available only to select trusted testers while Tavus works on safety measures.
Reported by Decrypt
What happened
AI startup Tavus said26 of 54 participants in its own test believed Griffin-Lite was a real person after a one-minute video call, a 48% rate against one believer out of 41 for its previous system, Decrypt reported. The company calls Griffin its first Human Interaction Model, built to handle face-to-face conversation with attention to expressions and pauses, and says it ranks first on Nvidia's VideoFDB benchmark with a generation score of 3.83 out of 5 against 2.80 for the next-best system and 3.92 for a human reference. Decrypt notes the figures come from Tavus' own research page, that participants were told they would be matched with another person, and that only at the end of the call were they asked whether the partner might not be real, so the setup differs from the classic Turing test proposed in 1950 by British mathematician Alan Turing. Griffin-Lite is limited to select trusted testers while Tavus works on safety measures. Only Decrypt covered the claim, and no independent verification of the test results appears in the evidence.
How the story developed
- Official statement
Tavus unveils the Griffin model
Tavus announced Griffin, described as its first Human Interaction Model, and said Griffin-Lite is limited to select trusted testers pending safety work.
- First report
First report by Decrypt
This AI Is Already Fooling People on Video Calls Into Thinking Its Human, Company Says
- Update
Decrypt reports the Tavus test results and benchmark claim
Decrypt published figures for the 54-participant Griffin-Lite test, the 41-participant prior system test and the Nvidia VideoFDB ranking, drawn from Tavus' own research page.
Affected entities
Original sources
| Publisher | Report | Role | Published |
|---|---|---|---|
| Decrypt Crypto media | This AI Is Already Fooling People on Video Calls Into Thinking Its Human, Company Says | Original |