A new forensics framework moves past asking whether a video is fake to identifying which system made it.
Detecting a synthetic fake has become the easier half of the problem. The harder question, which AI system generated a specific video, has largely gone unanswered, leaving investigators with confirmed fakes and no trail to follow. A research team at the University of California, Riverside is trying to close that gap.
Working with YouTube and Google's AI research division, the team has built a framework called SAGA, short for Source Attribution of Generative AI Videos. Instead of stopping at a real-or-synthetic verdict, SAGA classifies synthetic video clips by the specific text-to-video or image-to-video system that produced them. UC Riverside published details on July 24, 2026.
Knowing a video is AI-produced tells a platform or regulator that something is wrong. Knowing it came from a particular production system tells them where to look next: which model version, which access point, potentially which account or API key. That is the shift SAGA is built around, from detection toward attribution at the model level, a more precise instrument than a generic synthetic-content label.
Current detection tools were built when the question was whether a video was real. The spread of capable text-to-video and image-to-video systems has made provenance the more actionable question for both regulatory compliance and platform enforcement. Regulatory frameworks in multiple jurisdictions are beginning to require transparency about AI-produced content, and model-level attribution meets that demand more precisely than a generic label.
Google's AI research division is itself a developer of generative video systems, so the collaboration spans both the creation and the forensic identification of AI-generated content. The available materials do not characterize how that dual position shaped the research design, nor whether SAGA has been deployed in any production environment.
The framework's immediate test will be whether it generalizes across the expanding range of video production systems, not just the ones in its training data. A tool that identifies known generators accurately but fails on newer or less common systems would lose usefulness as the field keeps shifting. UC Riverside's published research is the next artifact available for independent scrutiny.
Get weekly leaderboard changes and AI video model launches delivered to your inbox.