There is no single standard. Some generators sign their output with a C2PA manifest, the format behind Content Credentials. Others write nothing at all and only leave a fingerprint in the filename or in a prompt string tucked into the container. A third group adds an invisible watermark to the pixels or the audio itself, which is a completely different mechanism and behaves completely differently when you edit the file.
The four carriers
In practice you will meet four carriers. A C2PA manifest sits in a dedicated segment or box and describes who made the file and with which tool. An XMP block holds fields such as CreatorTool and the IPTC DigitalSourceType flag. EXIF tags, PNG text chunks and ID3 tags are the oldest carriers and hold whatever the exporting software felt like writing, from the full prompt to the software version. All four are ordinary bytes and all four disappear when a file is re-encoded.
Why metadata is weak evidence
Because these carriers are ordinary bytes, they survive only if nothing touches the file. A screenshot loses everything. A messaging app that recompresses images loses everything. Uploading to most social platforms loses everything. That cuts both ways: a file with no Content Credentials tells you nothing about whether it was AI generated, and a file with credentials only tells you what the signer claimed.
Watermarks are not metadata
Google's SynthID and the audio watermarks used by several music generators live in the pixel or waveform data, not in a metadata block. They are designed to survive cropping, recompression and format conversion. No metadata tool removes them, ours included. If a page in this reference lists an invisible watermark, treat metadata removal as a privacy measure, not as a way to make the file untraceable.