University of California Riverside unveils SAGA tool to trace AI‑generated deepfake videos
Researchers at the University of California, Riverside have created SAGA, a video‑analysis system that can not only tell whether a video is real or fabricated but also attribute it to the specific generative‑AI model, version, and development team that produced it. The tool extracts spatial details from each frame and temporal patterns across frames—using a technique called Temporal Attention Signatures (T‑Sig)—to generate a multi‑level attribution.
SAGA was evaluated on public datasets that included videos generated by 19 different AI systems, covering both text‑to‑video and image‑to‑video models. The findings were posted on arXiv in 2026 under the title “SAGA: Source Attribution of Generative AI Videos.” By revealing the provenance of synthetic video content, the system is intended to bolster defenses against audiovisual misinformation, fraud and political manipulation.
The development was led by Rohit Kundu and collaborators at the university, and the work has drawn attention from media outlets in several countries as the first tool to provide detailed source attribution for AI‑generated deepfakes.
Entities
Generative AI video models · Rohit Kundu · SAGA · Temporal Attention Signatures · University of California, Riverside
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 4 SOURCES] SAGA uses a video‑transformer architecture with Temporal Attention Signatures to analyse spatial and temporal patterns. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] The tool was evaluated on public datasets that included videos generated by 19 different AI systems. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] The research describing SAGA was published on arXiv in 2026 with DOI 10.48550/arXiv.2511.12834. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] Researchers at the University of California, Riverside developed the SAGA tool. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] SAGA can attribute AI‑generated videos to the specific generative model, version, and developer that produced them. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] SAGA is intended to help combat audiovisual misinformation, fraud, and political manipulation. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] In testing, SAGA was evaluated on public datasets containing videos produced by 19 different generative AI systems, including both text‑to‑video and image‑to‑video models. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] A team at the University of California, Riverside developed SAGA, a tool that can attribute AI-generated videos to specific generative systems. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] SAGA uses a video transformer architecture with a robust vision model and Temporal Attention Signatures (T‑Sig) to detect distinctive temporal patterns. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] SAGA identifies whether a video is authentic or fabricated, the source modality, the AI model version, the development team, and the concrete generator. www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com
- [● 4 SOURCES] The research describing SAGA was published on arXiv in 2026 under the title “SAGA: Source Attribution of Generative AI Videos.” www.lne.es · www.farodevigo.es · www.laopinioncoruna.es · www.levante-emv.com