Glossary
Watermarking
Embedding an imperceptible signal inside media — increasingly inside AI-generated output — so its origin can be identified later. Unlike provenance metadata, the mark travels inside the file itself.
Approaches include imperceptible marks in pixels or audio and statistical watermarks in generated text, such as Google’s SynthID. Robustness is the open problem: marks must survive compression, cropping, and re-encoding, and research keeps showing removal is possible. Rules such as the EU’s AI transparency requirements are pushing labelling of AI output toward standard practice.
Why it matters
Watermarking helps trace media and label model output, but it complements contracts and provenance rather than replacing them.