New Forensic Tool SAGA Traces AI Videos to Their Source

A new forensic framework called SAGA can identify the specific AI system that generated a synthetic video, marking a shift from detection to source attribution. Researchers at UC Riverside, in collaboration with YouTube and Google DeepMind, published the work at CVPR 2026.

Beyond Detection

SAGA (Source Attribution of Generative AI Videos) analyzes unintentional visual patterns and temporal artifacts left by different generative models. The technique, called Temporal Attention Signatures (T-Sigs), visualizes unique patterns associated with each video generator.

"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," said Rohit Kundu, a UCR doctoral student who led the research under Professor Amit Roy-Chowdhury. "It turns out that, yes, there are distinct fingerprints there."

Multi-Level Attribution

The framework was tested against datasets containing videos from 19 different AI video generators, including both text-to-video and image-to-video systems. SAGA provides attribution across five levels: authenticity (real or synthetic), generation task (text or image source), model version, development team, and precise generator.

"Knowing that a video is fake is often not enough," the SAGA Research Team stated. "The critical need has shifted from whether it's fake to what is its source?"

Why It Matters

As synthetic videos become more realistic, identifying the specific generator allows investigators to track the origins of misinformation campaigns and helps regulators enforce transparency requirements for AI-generated content.

By pinpointing which platforms or models are being leveraged for deceptive activities, SAGA enables a more granular understanding of how synthetic media spreads. The paper is available on arXiv (arXiv:2511.12834).