How courts can prove AI-generated content is real

Deepfakes are now cheap and convincing, but courts and military commanders need reliable ways to prove whether AI generated something. This work bridges cryptography, watermarking, and law by mapping technical detection methods to the specific evidence standards that lawyers actually use. They release a 72,000-sample benchmark testing watermarking robustness against six attack types, evaluate four schemes, and provide concrete thresholds for when evidence is legally sufficient under armed-conflict rules, criminal procedure, and EU AI regulation.