AI Watermarking
Also called Content credentials, C2PA, Content provenance
Marking AI-generated images, video, audio or text, visibly or invisibly, so people and software can tell it was made by AI.
What is AI watermarking?
AI watermarking is any method of labelling content as AI-generated in a way that can be checked later. It's one of the main tools for telling real media from synthetic media such as deepfakes.
How does it work?
There are three main approaches, often used together:
- Visible marks: a logo or label on the image or video. Easy to see, and easy to crop out.
- Invisible watermarks: tiny changes to pixels, sound or word choices that people can't notice but a detector can. Google DeepMind's SynthID is a well-known example.
- Content credentials: signed information attached to a file recording who or what created it and how it was edited. The C2PA standard is used by many camera makers, software companies and AI providers.
What is it commonly confused with?
- AI detectors. Detectors guess whether content looks AI-made, with no mark to read, and they're often wrong. Watermarks are deliberately added by the tool that made the content, so checking them is more reliable, but only for tools that add them.
- Proof. Missing credentials don't prove content is real. Metadata can be stripped by re-saving or screenshotting, and invisible marks can be weakened by heavy editing.
Why does it matter?
As AI images, voices and video become hard to tell from real ones, watermarks and provenance help platforms, newsrooms and the public trace where content came from. Laws in several places, including the EU AI Act, now require some AI-generated content to be marked or disclosed. When you choose a media generator, check whether it adds watermarks or content credentials, and what its terms say about labelling.
