How to remove the Claude watermark

This page answers the question directly rather than burying it. If you only read one section, read the first.

The text watermark: no

There is no tool, paid or free, that removes the watermark from Claude-generated text, and the reason is structural rather than a matter of anyone building a better one.

Anthropic does not attach anything to the text. It biases the model’s word choices during generation so that the output carries a statistical signature spread across the whole passage. The watermark is not in the words in the sense of being a payload you could locate — it is the pattern of which words were selected.

So there is no character to strip, no metadata field to blank, no header to rewrite. A find-and-delete operation has nothing to operate on.

What the removers on the market do

They remove invisible Unicode characters — zero-width spaces, joiners, narrow no-break spaces, occasionally homoglyphs. You can watch this happen: paste marked text into the scanner on this site, which does the same operation, and you will see exactly what gets taken out. It is usually a handful of characters, sometimes none at all.

Those characters were never the watermark. They arrive from copying out of a rendered interface, the same way they arrive when you copy out of a PDF or a Word document. Removing them changes nothing about detectability.

Why the market persists anyway

Because nobody can check. Anthropic has not published the detection method, so no third party can test whether text still carries the mark. A removal tool cannot be shown to have failed, the buyer gets visibly “cleaned” text, and the product survives on the absence of any way to falsify it. If detection were public, that market would collapse within a week.

The documented limits, which are not a method

Anthropic publishes several conditions under which no signal is detectable: output from models predating the August 2026 rollout, passages too short to carry a signature, and heavily paraphrased text. These are stated as accuracy limitations of the system.

They are worth knowing because they cut both ways, and the direction that matters is the one nobody mentions: an absent mark proves nothing about whether content is AI-generated. Anyone treating a clean result as a clean bill of health has misread the tool. Rewriting your own text heavily enough to degrade a statistical signature also means you have rewritten your text heavily — at which point the question of whose words they are has changed anyway.

The C2PA manifest on images: yes, trivially

Files work completely differently. Claude attaches a C2PA manifest to supported image types — signed metadata recording what produced the file. This is an open standard, it is readable by anyone, and the inspector here reads it.

It is also extremely fragile. It does not survive a screenshot, most re-encoding, most resizing, or upload to the majority of social platforms. Metadata is stripped constantly, by accident, as a routine side effect of ordinary use.

That fragility is the real story about image provenance, and it is not good news for anyone. It means the signal is missing from most images in circulation regardless of origin, so its absence carries no information at all. The people it fails hardest are the ones who wanted provenance to work.

What actually helps

Sorted by how often the underlying problem is the one people really have:

• You were flagged and did not use AI. This is the most common real situation, and the watermark is usually not even what flagged you — commercial detectors are. What to do about it is a different problem with a real answer.
• You used Claude to proofread or translate your own writing. The mark is expected and its presence is not in dispute. Anthropic states that a detection means content may have been processed by Claude, not authored by it. Your argument is about what assistance your institution permits, not about the technology.
• Your text has invisible characters breaking something. Genuinely fixable, genuinely worth fixing, and the scanner does it — imports, keys, deduplication, search, diffs.
• You want to know what an image claims about itself. Read its manifest.

The one thing none of these is: a way to make AI-generated text pass as human. That product is being sold, it does not work, and the sellers are as unable to verify it as you are.