New Tool Traces AI-Generated Videos Back to Their Source

**SAGA (Source Attribution of Generative AI Videos)** is a research tool presented at CVPR 2026 that identifies which generative model produced an AI video by detecting unintentional visual artefacts as model 'fingerprints'; it performs multi-granular attribution (authenticity, generation task, model version, development team, specific generator), was trained on the DeMamba dataset covering 19 generators, and achieves strong results using only 0.5% labeled data per generator — enabling scalable forensic attribution for moderation, campaign tracking, and regulatory enforcement.