Can you remove the Claude watermark?

It depends which mark. C2PA provenance metadata on files can be stripped completely and verifiably. The statistical text watermark can only be weakened by substantial rewriting, and because Anthropic has not released a detector, nobody can currently verify the result.

Does Claude use invisible Unicode characters?

No. Anthropic describes a watermark woven into the text itself, which technical reporting describes as biased token selection. Zero-width characters would be trivially removable and therefore useless as a provenance signal.

When did Claude start watermarking?

Models launched on or after 2 August 2026 support machine-readable marking at launch. It covers Claude, the API, Claude Code, Claude Cowork and Claude Tag, worldwide, with no opt-out.

Does the watermark prove Claude wrote something?

No. Anthropic states that a detected mark means content may have been processed by Claude and is not fully conclusive. It does not establish authorship, which matters if you used Claude only to proofread your own writing.

Marking live since 2 August 2026

Remove Claude Watermark
the right way.

Remove Claude watermark marks correctly and the first thing you need is the one fact almost every tool in this category gets wrong: Claude does not hide invisible Unicode in your text.

Input

0 words

Output

Your cleaned text will appear here.

Processed in your browser

The distinction

The mark is in the words, not between them.

The text mark is a statistical pattern in word choice, and files carry signed C2PA metadata instead — two different marks, and neither is removed by the character scrubbers being sold as the answer.

Most removers target the wrong thingThey strip zero-width characters. Claude doesn’t use them.

A document with several word choices highlighted, beside a panel reporting a detected provenance mark

The three marks

Three marks get confused as one

Almost every argument about this topic dissolves once you separate them.

Mark 1

Text watermark

Statistical bias in token selection, keyed to a secret

Removable: Only by heavy rewriting
Verifiable: No — no public detector

Mark 2

C2PA file metadata

Signed provenance manifest in .png, .jpg, .svg

Removable: Yes, completely
Verifiable: Yes — read the file

Mark 3

Invisible Unicode

Zero-width characters. Not used by Claude

Removable: Yes, trivially
Verifiable: Yes — but irrelevant here

Start here

The pages people open first

All of it

Browse by what you need

FAQ

Common questions

Can you remove the Claude watermark?: It depends which mark. C2PA provenance metadata on files can be stripped completely and verifiably. The statistical text watermark can only be weakened by substantial rewriting, and because Anthropic has not released a detector, nobody can currently verify the result.
Does Claude use invisible Unicode characters?: No. Anthropic describes a watermark woven into the text itself, which technical reporting describes as biased token selection. Zero-width characters would be trivially removable and therefore useless as a provenance signal.
When did Claude start watermarking?: Models launched on or after 2 August 2026 support machine-readable marking at launch. It covers Claude, the API, Claude Code, Claude Cowork and Claude Tag, worldwide, with no opt-out.
Does the watermark prove Claude wrote something?: No. Anthropic states that a detected mark means content may have been processed by Claude and is not fully conclusive. It does not establish authorship, which matters if you used Claude only to proofread your own writing.

Source

Everything here is checked against primary documentation rather than other tools’ marketing copy.

Anthropic: How Claude marks AI-generated content →

Remove Claude Watermarkthe right way.