Researchers at the University of California, Riverside have built a tool that goes past spotting a fake video. It names the AI system that made it.
SAGA
The framework is called SAGA, short for Source Attribution of Generative AI Videos, and it represents one of the first large-scale efforts to trace AI-generated videos back to their origins. It works by reading visual patterns buried in the frames, artifacts that video generators leave behind without meaning to. Those accidents turn out to be a forensic trail.
Fingerprints
UCR doctoral student Rohit Kundu led the research under professor Amit Roy-Chowdhury, working with researchers at YouTube and Google DeepMind. “The patterns are like fingerprints that the generative model leaves behind, and our goal here was to find out if the signatures are distinct amongst different generators. It turns out that, yes, there are distinct fingerprints there,” Kundu said.
Temporal Attention Signatures
A still image gives a detector one surface to inspect. Video gives it motion, and motion is where generators give themselves away. The team found that different systems introduce their own subtle irregularities in how visual information shifts between frames.
SAGA reads that change over time rather than picking apart a single frame. The result is what the researchers call Temporal Attention Signatures, or T-Sigs, which make the fingerprints visible and hold up even against generators the system hasn’t seen before.
Characteristic profile
Averaging the patterns from videos made by the same system produces a profile that separates one generator from the next. The researchers tested SAGA on public datasets covering 19 AI video generators, spanning both text-to-video and image-to-video systems.
The attribution runs five levels deep: whether the video is real, whether it came from text or an image, the model version, the development team behind it and the specific generator that produced it.
The efficiency is arguably the bigger result. SAGA matched fully supervised performance using just 0.5% of the source-labeled data per class, which matters because labeled training data for every new generator is exactly what nobody has time to assemble.
What we think
Detection tools that flag a video as synthetic keep arriving. Knowing a clip is fake was the useful question two years ago. It isn’t anymore, not when the fake is engineered to move votes or money, and not when the follow-up question is who built the thing that made it.
For anyone producing video for a living, attribution cuts both ways. It gives platforms a route to enforcement and gives you a route to proving your own footage came from a camera. Both matter as licensing disputes and provenance rules tighten.
Worth keeping expectations calibrated. SAGA is published research, presented at the CVPR 2026 conference, not a product you can point at a suspect file today. What it demonstrates is that the trail exists. Someone still has to build the road.
