In plain English
Watermarking builds an invisible signal into what a model produces, so that software can later recognise the content as machine-generated. In images it is a pattern of pixel changes the eye cannot see. In text it is a subtle statistical bias in which words get chosen. Related provenance standards instead attach signed metadata recording how a file was made and what edited it since.
Why it matters
As synthetic media becomes indistinguishable by eye, provenance is one of the few practical defences left, and several jurisdictions are moving towards requiring it.
A worked example
An image tool marks every picture it generates, so a platform can flag the file as AI-generated even after it has been reposted without any caption.
Common confusion
Watermarking is not detection. It only works on content that was marked at the point of creation, so unmarked or deliberately stripped output will sail straight past it.