Tavus, a small AI startup based in San Francisco, has launched Griffin, a model built to hold a live video conversation the way a person would. It watches your face, listens to your tone, notices pauses, and responds on camera in real time instead of just typing back text.
In the company's own test, 48 percent of people who spoke with Griffin for one minute over video believed they were talking to a real person. The previous best AI system only fooled 2 percent of people. That is a different category of result, and it shows the gap between an AI that simply answers and one that feels present on screen has closed fast this year.
A second test, which Tavus says was run independently by Nvidia, rated how human an AI feels on a five point scale during a live video chat. Griffin scored 3.83 on that scale.
A real human scored 3.92, only slightly higher. The best AI system before Griffin managed just 2.80. Scores that close to a real person matter because this kind of live, two way video test is usually where AI systems fall apart, not where they shine.
Keep in mind these are numbers Tavus itself is publishing, with Nvidia named as the outside grader, so treat the figures as a strong claim from the company rather than an independently verified result. Still, the direction is hard to miss: video AI is getting harder to tell apart from a real person on a short call.
Tavus was founded in 2020 and has raised 64 million dollars across two funding rounds, an 18 million dollar round in 2024 and a 40 million dollar round this past November. It started by making pre recorded personalized videos for sales and marketing teams. It has since moved into live, two way video conversations with AI personas, a much harder problem and a much bigger market if it works.
For now, Griffin is only available as a limited preview called Griffin-Lite, given to a small group of testers. Tavus says a fuller version is coming once it works through safety concerns, and lists tutoring, practicing hard conversations, and camera based tech support as the main uses it has in mind. A sales rep could rehearse a tough client pitch, a student could get one on one tutoring on demand, or a customer could be walked through a broken printer by a patient voice on screen that never gets tired.
The flip side is harder to ignore. Video call fraud is already a real cost to businesses. A finance worker in Hong Kong approved a transfer of 25.6 million dollars after a video call with people who turned out to be deepfakes of his own company's executives.
Fraud trackers at Pindrop reported deepfake fraud attempts jumped 1,300 percent in 2024 alone, and fake candidates using real time face swaps are now showing up in job interviews too. A model built specifically to pass as human on a live call makes that kind of fraud easier to pull off for anyone who wants to misuse the technology, not just Tavus's own customers.
The practical lesson is simple. A face and a voice on a video call are no longer proof of who you are talking to. If your company still approves wire transfers, contract changes, or hiring decisions based on a video call alone, this is the moment to add a second way to confirm identity, something a screen cannot fake yet, like a callback to a known phone number or a code word agreed on in advance.