Nvidia Built a Tool That Tries to Spot AI-Generated Video in Milliseconds
Nvidia's new Synthetic Video Detector claims up to 92% accuracy at flagging AI-made video, fast enough for live broadcasts. A genuinely useful idea, with limits worth understanding before you rely on it.
Nvidia has released a tool aimed at one of the more unsettling side effects of the AI boom: video you cannot trust. Called the Synthetic Video Detector and shown at the SIGGRAPH graphics conference, it tries to tell whether a clip was generated by AI, and it does so fast, processing 1080p video in as little as 22 milliseconds. Nvidia reports up to 92% accuracy on clean, uncompressed video. There is a demo you can try today at build.nvidia.com, though the free cloud version is slow, caps files at 100MB, and often times out.
Here is the clever part, in plain terms. Most people assume you catch a fake by spotting obvious mistakes: six fingers, warped text, a face that flickers. This tool does not look for those. Instead it hunts for the statistical fingerprints that AI video generators leave behind, subtle patterns in the pixels that come from how these models build a frame, invisible to your eye but detectable by another model. Technically it uses a “vision transformer,” a type of AI designed to analyse images, built on top of well-known open models called DINOv2 and DINOv3. You do not need those names to get the point: it is AI used to police AI.
That framing is the whole story. As tools like Sora and its rivals make convincing video trivial to produce, the defence is not a human squinting at a screen but automated detectors running quietly in the background. Nvidia plans to ship this one through a streaming company’s toolkit so newsrooms and platforms can score live video for a “synthetic” probability, on-premises or even in air-gapped setups that never touch the internet.
What this means for you: Right now this is aimed at broadcasters and platforms, not at you personally, so do not expect a “fake or real” button in your video app tomorrow. But the direction is reassuring: the people who move video around at scale are getting instruments to flag manipulated clips before they spread. The caveat is important, though, and Nvidia is upfront about it. Accuracy drops as video gets compressed, falling to about 87% at light compression and 82% at heavy compression, and almost everything you watch on social media is heavily compressed. So a detector like this is a helpful signal, not a verdict. The safest habit remains the old one: treat a shocking clip from an unknown source as unconfirmed until a name you trust stands behind it.
Sources
Source: https://build.nvidia.com/nvidia/synthetic-video-detector
Nvidia's Next AI Chips Reportedly Squeeze 10x More Work From the Same Power
CoreWeave says Nvidia's new Vera Rubin systems produced about 10 times more tokens per megawatt than last generation's Blackwell in an early benchmark. Here is what 'tokens per megawatt' means and why it decides whether AI can keep growing.
