Remove
Remove Claude AI Watermark
Remove Claude AI watermark requests almost always mean one of two completely different marks, and the fix for one does nothing for the other.
Claude applies a statistical watermark to text and signed C2PA provenance metadata to files. Identify which you are holding before you try anything.
File metadata strips cleanly and verifiably. Text marks only weaken under heavy rewriting.
Runs in your browser. No signup, no upload, nothing stored.
Two marks, not one
Anthropic marks text and files by different mechanisms, because the two formats allow completely different things.
Text gets a watermark woven into the writing itself. Files — specifically .svg, .png and .jpg — get signed provenance metadata following the C2PA open standard.
- Text: statistical, invisible, no character to delete, no public detector
- Files: metadata blocks in the container, fully inspectable, removable today
If you have a file, this is straightforward
C2PA data lives in defined places: a JUMBF box behind the APP11 marker in JPEG, a dedicated chunk or XMP packet in PNG, an XML metadata block in SVG.
Stripping those blocks removes the provenance claim without touching a single pixel. Re-encoding the image, converting the format or taking a screenshot also destroys it, but each of those degrades the file, which byte-level stripping does not.
Best of all, you can confirm the result. Read the metadata again and it is either there or it is not.
If you have text, expectations have to drop
There is no equivalent clean operation, because there is no discrete thing to delete. The mark is spread across the word choices in the passage.
Anthropic lists heavy editing, paraphrasing, translation and short length as conditions under which a mark may go undetected. Those are the only levers anyone has, and none of them can be verified while detection tooling remains unreleased.
Why the distinction keeps getting lost
Most tools in this space were built for the previous generation of AI-detection folklore, when the story was that chatbots inserted hidden Unicode as trackers.
They kept the marketing and added Claude to the headline. The mechanism never changed, so the tool addresses neither of the two marks Claude actually applies — and its users leave believing they have handled something they have not.
A quick way to tell which you have
If you can select the words with a cursor, it is text and the statistical mark applies. If you are holding a file you would attach to an email, check the metadata first.
Mixed cases are common and worth splitting deliberately. A PDF export of Claude-written prose carries both: document metadata in the container, and the statistical mark in the sentences inside it.
Source
The claims on this page are drawn from primary documentation and reporting rather than from other tools’ marketing copy.
Anthropic: How Claude marks AI-generated content →Related
Check what your text and files actually carry
Runs in your browser. No signup, no upload, nothing stored.