Can you remove Anthropic's text watermark?

An honest answer: what degrades a statistical watermark, what cannot touch it, and why nobody can currently verify the result.

Anthropic's watermark · 3 min read

Since 11 August 2026 Claude embeds an invisible watermark in generated text. The obvious next question is whether it can be removed. Here is the accurate answer, including the parts that are inconvenient for a service that does this.

What cannot remove it

Deleting hidden characters. Stripping zero-width characters, Unicode tag characters and variation selectors is exact and verifiable — and irrelevant here. A statistical watermark is not a character sitting in your text. It is a pattern in which words were chosen. There is nothing to delete.

This matters because plenty of "watermark remover" tools do only character cleanup. Against a statistical mark, that accomplishes nothing.

Normalising punctuation. Same reasoning. Converting em dashes to hyphens and curly quotes to straight ones removes real stylistic tells. It does not touch a token-selection pattern.

Retyping it yourself. This actually would work, because you would be choosing the words. It is also slower than writing it yourself in the first place.

What Anthropic says degrades it

From Anthropic's own documentation, the mark may not survive:

• heavy editing
• paraphrasing
• translation
• mixing with other content
• very short passages

That list is the vendor's, not ours, and it is the most authoritative statement available. Paraphrasing appears on it. So does heavy editing — note "heavy", not "light".

This follows from how statistical watermarking works. The signal lives in the specific tokens selected among near-equivalent alternatives. Replace those tokens and the accumulated pattern degrades. Google documented the same weakness for SynthID-Text, and it is the reason OpenAI held its own watermark back rather than shipping it.

The part nobody can honestly promise

Anthropic has not published its detection mechanism. Its documentation says technical details are "forthcoming".

Until that exists, no one outside Anthropic can check whether a given passage carries the mark. That includes us. Which means:

• Nobody can demonstrate that removal succeeded.
• Nobody can demonstrate that it failed.
• Any service claiming verified removal of Anthropic's watermark is asserting something it has no way to test.

We would rather say that plainly than sell you a number we made up. When a detector becomes available we will test against it and publish what we find, including if the answer is unflattering.

What this service actually does

Two layers, and we are precise about which is which:

Layer 1 — exact, verifiable. Every zero-width character, Unicode tag character, variation selector, exotic space and typographic punctuation mark is removed by direct codepoint matching. Re-scan the output and the count is zero. No model involved, no interpretation, nothing can drift. This is real and provable, and it has nothing to do with the statistical watermark.

Layer 2 — a rewrite. The text is rewritten to clear vocabulary and cadence markers, changing as little as possible. This is a paraphrase pass, which is the mechanism Anthropic itself names as degrading the mark. Whether it degrades it enough is not something anyone can currently measure.

We will not tell you Layer 2 defeats the watermark. We will tell you it is a paraphrase, that paraphrasing is on Anthropic's list, and that the outcome is unverifiable today.

What the mark actually means for you

Worth keeping in proportion. Anthropic states a detected mark means content may have been processed by Claude — not that Claude authored it. Text you wrote yourself and asked Claude to proofread carries the same mark as text Claude drafted.

So a mark is not evidence of what people usually assume it is evidence of. If you are facing an accusation, that distinction is more useful than any removal tool: the mark cannot distinguish "Claude wrote this" from "Claude edited my writing".

And if you have an institutional obligation to disclose AI assistance, no tool resolves that. Removing a mark does not change what you did.

Run a free scan to see the character layer exactly. On the statistical layer, we will keep telling you what is known.

Related: does Anthropic watermark Claude's text? · EU AI Act Article 50 · how detectors work

Sources: Anthropic Help Center