Why a visible AI watermark is not enough

A visible watermark is familiar: a small logo, “AI generated” text, or a pattern over the image. It may be useful as a notice for people, but it is not the same thing as machine-readable marking. It can be cropped, covered, copied onto another image or simply lost in a new export.

For the technical side of Article 50(2) of the EU AI Act, that difference matters. The provision calls for machine-readable marking for AI-generated or manipulated content. This is general technical information, not legal advice; a lawyer can assess a particular obligation. The practical point is that a label needs to travel as information a system can read, not only as pixels a person can see.

What a visible watermark can and cannot do

A visible watermark can be a clear disclosure to a human viewer. It may make sense in a campaign, a preview image or an editorial context. But the file itself cannot use it to prove which tool made the image, whether the notice was added later, or what edits happened between creation and delivery.

• It is part of the picture and can be removed by cropping or retouching.
• It does not identify a trustworthy source by itself.
• It cannot reliably distinguish the original file from a copy.
• It gives automated systems no standard way to inspect provenance.

What C2PA Content Credentials add

C2PA Content Credentials are a cryptographically bound provenance record associated with a file. Depending on the workflow, they can record that a file was created with an AI tool, that it was a digital capture, and which edits followed. The important difference is verification: changing the file can invalidate the connection to the record.

That does not make the mark magical. It does not prove every claim about an image, and it does not decide after the fact whether an unmarked file was AI-made. It gives you a concrete, checkable record for files that were marked in the workflow from the start.

Why “we put a watermark on it” is not the end of the job

Even a correctly marked source file may move through a designer’s export, a web shop, a CDN, a social platform or a customer download. Those normal steps can create a new file. A visible watermark may still be seen, while the cryptographic provenance record is missing; or the reverse may be true in a workflow that does not display a visible notice.

So the useful question is not simply whether there is a mark on the image. It is: what does the file contain at origin, and what reaches the user after the actual delivery path? Our guide on C2PA label survival on social media shows common break points.

A simple check for your own output

1. Take one file straight from the tool that creates it.
2. Check whether it contains Content Credentials.
3. Send it through one real route you use: design export, website, customer download or platform upload.
4. Check the resulting file too, and record the difference.

That small test is more useful than guessing from a logo in the corner. It shows where your workflow needs technical attention and where the result can be demonstrated.

AI Detective helps put machine-readable marking into the workflow that produces your files, then maps important delivery paths. We do not sell an AI detector and we do not give legal advice.