Claude watermark in code: what it means for your repository

Official sources cited Answer in the first paragraph No upload

Code is text, so the same rules apply. Anthropic describes watermarking as applied to its text output and does not carve out source code. But three things follow from that which matter more than the label: nothing is added to your files, the code still does exactly what it did, and no cleaner — including this one — can strip the watermark out of a .py or .js file.

What is real, and what does break builds, is a completely different problem: invisible characters that arrive in code through copy-paste and survive into your repository. Those are bytes in your files. You can find them, you can delete them, and you can stop them coming back.

TWO DIFFERENT THINGS IN A REPO Only one of them breaks anything. statistical watermark in the token choices not removable · no public checker does not change behaviour invisible characters real bytes in the file removable · scannable breaks diffs, linters, builds

What the watermark does not change

It is worth being precise here, because most of the alarm about code is aimed at the wrong target.

Which leaves the question people actually care about: can someone else tell? That is answered on the detector page — the short version is that no public tool can confirm it, because reading it needs Anthropic's key.

What does break builds

Invisible characters. These arrive in code exactly the way they arrive in prose: copying from a chat interface, from a web page, from a PDF, or out of an editor that inserted them. A zero-width space inside an identifier is not a syntax error your eyes will catch, because it has no width. It is a different character to the compiler.

The symptoms are familiar once you know to look for them: a line that fails to match in a diff for no visible reason, a linter complaining about an invisible token, a string comparison that should pass and does not, a YAML or JSON file that parses locally and fails in CI.

This part is fixable. The folder guide has a working script with a dry-run mode, plus a pre-commit hook so they cannot come back into the repository. Run the dry run first and read what it would change before it writes anything.

Provenance in a repository

Anthropic attaches a C2PA content credential to files it produces, and the formats named are images and documents — PNG, JPG, SVG. Plain source files are not in that list, so a .py or .ts file Claude wrote does not normally carry one. If you are handling generated images or exported documents in a repository, that is a different matter, and that credential is removable.

Separately: git records who committed and when, not how the contents were produced. There is no watermark field in a commit object. If your organisation needs to record AI assistance, that is a policy decision expressed in commit messages or a trailer — not something the tooling does for you.

The honest limits

We cannot remove Claude's statistical watermark from code or from prose, and no other tool can either — it is a property of how the tokens were chosen, not of the bytes in your file. Any site selling a "Claude watermark remover for code" is either rewriting your source or selling you nothing. What we can do is clean the layer that is real bytes and does cause actual breakage, and tell you plainly which of the two you are looking at.

Frequently asked questions

Does the watermark apply to code Claude writes?

Anthropic describes watermarking as applied to its text output and does not carve source code out, so treat code as covered. That said, we are reporting what the documentation covers rather than quoting a sentence about code specifically — if your compliance question turns on that distinction, read the source document itself. What is not in doubt is that nothing is added to your files either way.

Can I strip the watermark out of my source files?

No, and neither can anything else. It is not stored in the file. Deleting comments, reformatting, or running the code through a cleaner changes the bytes and leaves the pattern alone, because the pattern is in which tokens were chosen in the first place. Rewriting the code yourself is a different act — that produces different code — but that is rewriting, not stripping.

Is watermarking source code a software sovereignty issue?

That argument has been made, and the honest answer is that it is a policy question rather than a technical one. The technical facts are narrow: nothing is added to your files, the code behaves identically, and detection needs a key that is not public. Whether a jurisdiction treating model-generated code as subject to marking requirements creates a dependency is a matter for the people writing those rules, not something a cleaner can resolve.

Will the watermark break my build or my linter?

No — it is not in the file, so nothing reads it during a build. If your build is breaking on something invisible, it is almost certainly a zero-width or directional character that arrived through copy-paste, which is a different problem and one you can actually fix. Run the folder script with its dry-run mode first and see what it finds.

Does git record anything about how the code was produced?

No. A commit records the author, the committer, the timestamps, the message and the tree — there is no field for how the contents were generated, and no watermark in a commit object. If your team needs to declare AI assistance, do it in the commit message or a trailer, because nothing in git will do it for you.