How to remove Claude's watermark from text

A step-by-step guide to removing Claude's watermark: what can be stripped exactly, what needs rewriting, and what nobody can verify yet.

Removing the watermark · 3 min read

Since 11 August 2026, Claude embeds an invisible watermark in text it generates. If you want to remove Claude's watermark, this is the accurate procedure — including the part where nobody can currently prove it worked.

What you are actually removing

There are two separate layers, and confusing them is why most advice on how to remove a Claude watermark is useless.

Layer 1: hidden characters. Zero-width characters, Unicode tag characters, variation selectors, exotic spaces, and typographic punctuation like em dashes and curly quotes. These are real codepoints sitting in your text. They can be found by exact matching and deleted losslessly, and you can verify the result by re-scanning.

Layer 2: the statistical watermark. A pattern in which words the model chose. There is nothing to delete — the signal lives in the word selection itself. Only rewriting touches it.

Any tool that only does Layer 1 and calls itself a Claude watermark remover is not addressing the watermark at all.

Step 1: paste, never retype

Copy the text from Claude and paste it. Retyping destroys the evidence you are trying to inspect — every hidden character survives copy-paste and none survives being typed by hand.

Step 2: scan it

Run a character-level scan first, so you know what you are dealing with before you change anything. A scan gives exact counts per category, not a probability. Claude output reliably contains:

• em dashes, used heavily as a mid-sentence connector
• curly quotes and apostrophes rather than straight ASCII
• the single-codepoint ellipsis U+2026
• occasionally zero-width characters and non-breaking spaces

Our scan is free, unlimited, and runs in your browser — the text never reaches a server.

Step 3: strip the characters

Delete every invisible codepoint and normalise the punctuation to ASCII. This is deterministic: no model, no interpretation, nothing about your wording can drift.

Verify by re-scanning. The removable count should be zero. This is the only part of Claude watermark removal that is provable.

Step 4: rewrite the wording

Here is what Anthropic's own documentation says the mark may not survive:

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

Paraphrasing is on Anthropic's list. That is the mechanism, and it comes from the vendor, not from us. So the rewriting step is not a guess about what might work — it is the thing Anthropic itself identifies as degrading the mark.

Practically, that means:

• Replace signature vocabulary: delve, leverage, underscore, robust, seamless, tapestry, realm, pivotal
• Delete scaffolding connectives: sentence-opening Moreover, Furthermore, Additionally, In conclusion
• Cut empty hedges: it is important to note that
• Vary sentence rhythm — model prose runs sentence after sentence of near-identical length

Note the word "heavy" on Anthropic's list. A one-word-per-paragraph touch-up is not what they are describing.

Step 5: accept what cannot be verified

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

Until that exists, nobody outside Anthropic can check whether a passage still carries the mark — not you, not us, not a university, not a publisher. So:

• No one can prove removal succeeded.
• No one can prove it failed.
• Any service claiming verified Claude watermark removal is asserting something it cannot test.

We would rather tell you that than invent a percentage.

What a detected mark actually means

Worth keeping in proportion, because people over-read it. Anthropic states a mark indicates 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 from scratch. The mark cannot distinguish those two cases. If you are facing an accusation, that distinction is more useful than any removal tool.

Quick reference

Run the free scan to see what is in your text right now.

Related: does Anthropic watermark Claude's text? · can it be removed? · how to check for it

Sources: Anthropic Help Center