AI video model Griffin fools 48 percent of users in Turing test
Tavus built a real-time AI video avatar that passes for human. Half of test users thought they were talking to a real person.
Models · Source: NY Post
What happened
Bay Area startup Tavus just revealed Griffin. They call it a Human Interaction Model. It generates real-time video avatars from a single photo. It renders every pixel of every frame live. That includes facial expressions, arm movements, chair shadows, and background shifts. It listens, watches expressions, and talks all at once. In demonstrations, it even taught a user how to solve a Rubik's Cube and played Simon Says.
The company ran a one-minute video call test with 54 people. The results are startling. Exactly 26 of those participants thought they were speaking to a real human. That is a 48 percent success rate for the machine. The users who got fooled were 79 percent confident in their guess. Those who correctly spotted the AI were 81 percent confident. People who figured it out usually did so in under 20 seconds.
This is a massive leap from older technology. Tavus notes their previous software stack only tricked one out of 41 people. Griffin reacts to visual cues, handles interruptions, and adjusts its tone on the fly. But you cannot buy it today. Tavus is keeping it restricted to trusted testers. They are delaying general availability to build safety guardrails and prevent deepfake fraud.
Key facts
- 48% — Percentage of test users who thought the AI was a real person
- 26 — Number of users out of 54 who were fooled by the model
- 79% — Average confidence of users who incorrectly thought the AI was human
- 3.83 — Griffin's score on the NVIDIA VideoFDB generation benchmark
- 2.4% — Pass rate of Tavus's previous AI model stack
Why it matters
The barrier to creating photorealistic video agents just collapsed. Builders can soon deploy AI that handles customer support, tutoring, and job screening without the robotic uncanny valley. You no longer need complex rendering pipelines. A single reference image and a strong model can now sustain a believable human conversation. This fundamentally alters how we design digital interactions. Builders can create applications where the interface is just a face on a screen. This reduces friction for non-technical users who struggle with traditional software menus.
Trust in video calls will plummet. If a startup can generate a convincing human in real time, bad actors will weaponize this for deepfake fraud. The multi-million dollar video-call fraud pattern is already a reality. This technology makes that kind of scam cheaper and easier to execute. We will see a massive surge in identity verification tools. Companies will have to prove their human employees are actually human. The default assumption on a cold video call will shift from trust to extreme skepticism. The entire remote work infrastructure relies on visual trust, and that trust is now obsolete.
For builders
Build high-touch automated video services
Customer support, telehealth, and tutoring just got a massive upgrade. You can build platforms where users interact naturally with video agents without feeling alienated. Companies paying for massive offshore human call centers stand to lose market share to automated AI solutions.
Prepare for identity verification demand
Fraudsters will absolutely use this technology to spoof job interviews and executive conference calls. There is an urgent opportunity to build real-time deepfake detection tools. Cybersecurity vendors and identity verification startups will make a fortune securing enterprise video pipelines.
Wait for general API availability
Griffin is currently locked behind a trusted tester program called Griffin-Lite. Do not rip out your existing video generation stack just yet. Tavus is intentionally delaying the public launch to build safe disclosure features and prevent malicious use by bad actors.
My take
I always tell founders to look at the second-order effects of new models. If an AI can fake a job interview this well today, our entire remote hiring process is fundamentally broken. We have to stop trusting our eyes on video calls and start building cryptographic proof of human identity.
Original reporting: NY Post. This is my rewrite and opinion.