Nvidia unveils AI tool that detects fake videos at rapid speed

Nvidia has unveiled an AI tool that can detect fake and manipulated videos in a fraction of a second, as the technology industry races to counter the spread of increasingly convincing deepfakes.

Called the Synthetic Video Detector (SVD), the tool was announced at the SIGGRAPH 2026 computer graphics conference. It analyses footage frame by frame and returns a probability score indicating how likely a clip is to have been generated or altered by AI — reportedly in as little as 22 milliseconds.

Nvidia says the detector achieves up to 92 percent accuracy on uncompressed video. That accuracy declines as footage is compressed, falling to around 87 percent at 15 percent compression and 82 percent at 50 percent a notable limitation given that platforms such as YouTube, TikTok and Instagram routinely compress uploads, stripping out the subtle artefacts that detection models rely on.

Rather than delivering a definitive verdict, the tool produces a classifier score that editorial and media teams can use to prioritise, flag, quarantine or escalate suspect footage. Nvidia has positioned it as an additional layer of confidence that complements, rather than replaces, traditional fact-checking and digital forensic analysis.

The detector is being offered as part of Nvidia's NIM microservices within its “AI for Media” platform, allowing newsrooms, broadcasters and enterprises to integrate video verification into their existing workflows. It is aimed at professional users rather than consumers and is not publicly available, though Nvidia has demonstrated it, and the company says it can assess 1080p video at scale.

According to Nvidia, the latest version tops an industry benchmark for synthetic-media detection, the AI GVD Bench, and is built on award-winning research presented at leading computer-vision conferences.

The launch comes as deepfakes from manipulated political speeches to fabricated news footage and AI-generated celebrity clips have evolved from an online curiosity into a serious challenge for journalism, cybersecurity and public trust.