The mechanism differs by medium: image and video generators typically embed a statistical pattern in pixel values (Google DeepMind's SynthID-Image is one example), while text watermarking re-derives the randomness behind a model's word choices rather than altering pixels or adding hidden characters — see AI Text Watermark for that mechanism specifically. Watermarking is distinct from C2PA content-credential metadata, which is a signed, human-readable manifest attached to a file rather than a pattern baked into the content itself; metadata is stripped by a screenshot or re-save, which is why labs increasingly ship both together. Regulatory pressure, especially the EU AI Act's Code of Practice on Transparency of AI-Generated Content, is the main driver behind major labs adopting watermarking through 2026.