As AI-generated video becomes indistinguishable from real footage to the naked eye, investigators and platform moderators have had one reliable question they could still ask: "Is this fake?" A new tool called SAGA (Source Attribution of Generative AI Videos) shifts the forensic frontier to a far more powerful question: "Who made it?"
Developed by doctoral researcher Rohit Kundu and Professor Amit K. Roy-Chowdhury at UC Riverside's RAISE Institute, in collaboration with researchers from YouTube and Google DeepMind, SAGA was presented at IEEE/CVF CVPR 2026 — one of the premier computer vision venues in the world.
How SAGA Works
Every AI video generation system leaves behind unintentional visual artefacts in the frames it produces. These patterns are not watermarks deliberately embedded by the developer — they are byproducts of the underlying generative architecture, training data, and rendering pipeline. SAGA is designed to detect, characterise, and compare these signatures.
"The patterns are like fingerprints that the generative model leaves behind, and our goal here was to find out if the signatures are distinct among different generators," said Kundu. "It turns out that, yes, there are distinct fingerprints there."
What makes SAGA particularly powerful is its multi-granular attribution — the tool does not just say "this is synthetic." It operates at five levels simultaneously:
- Authenticity — real video vs. AI-generated synthetic
- Generation task — text-to-video (T2V) vs. image-to-video (I2V)
- Model version — distinguishes between iterations of the same system
- Development team — identifies the organisation that built the generator
- Specific generator — names the precise model responsible
Coverage and Efficiency
SAGA was trained and evaluated on the DeMamba dataset, spanning 19 different AI video generators across both text-to-video and image-to-video modalities. The tool achieves state-of-the-art attribution performance using only 0.5% of source-labelled data per generator through a pretrain-and-attribute strategy — making it feasible to extend to new generators as they emerge, without retraining from scratch.
Why Attribution Matters More Than Detection
Detection tools have a fundamental limitation: they answer a binary question, and adversaries can train new generators specifically to evade known detection signatures. Attribution changes the game. Even if a synthetic video slips past a "fake or not" classifier, a forensic attribution tool can tie it to a specific platform or toolchain — enabling enforcement actions, tracking misinformation campaigns to their origin, and building evidence chains for regulators.
This is especially relevant in nation-state influence operations, where the same AI pipeline is often used across a coordinated campaign. Identifying the generator ties all videos in the campaign to a common source, even when content differs wildly.
Regulatory Context
The EU AI Act, which came into force in 2024 and began applying in stages through 2026, includes provisions requiring providers of AI systems that generate synthetic audio-visual content to implement technical measures enabling detection. SAGA represents the kind of capability that makes those provisions enforceable in practice — not just on paper.
The research is publicly available via arXiv and the CVPR 2026 proceedings. The team is working with YouTube and Google DeepMind on pathways to deployment for content moderation at scale.
As AI video generators proliferate and their output quality improves, tools like SAGA may become as fundamental to digital media forensics as hash-based file integrity checking is to traditional cybersecurity — a baseline capability that shifts the burden back onto the attacker.