removeclaudewatermarks.com
← removeclaudewatermarks.com

11 Aug 2026Announcement

Does Claude watermark text? What actually changed

Anthropic began marking Claude’s output on 11 August 2026. It covers every model and product, worldwide, and you cannot switch it off. Here is what is marked, what can be removed, and — more usefully — what Anthropic itself said about the limits.

01What is actually marked

Two separate mechanisms, which almost every write-up has merged into one. Text carries an invisible statistical watermark embedded at the model level, in the choice of words themselves. Files such as images carry digitally signed provenance metadata following the C2PA standard.

The distinction is not academic. It decides what can be removed and what cannot, and it is the reason a tool that strips hidden characters cannot do what its name suggests.

02What Anthropic conceded

Buried in the coverage is the most useful sentence for anyone trying to work out their options. Anthropic acknowledged that text watermarks may be erased if the writing is heavily edited, and that a file’s metadata “could disappear if the format of the file changes” — or if a screenshot is taken.

Read that as a specification. Changing a file’s format clears the image mark, which is a browser-side operation. Heavy editing is the only lever on the text mark, and heavy editing means replacing your words — not deleting invisible ones.

Anthropic also said the mark will “travel with the text when it’s copied and pasted elsewhere.”

Reported by Forbes, 11 August 2026

03Why people are annoyed

The backlash on Reddit and X was immediate, and the sharpest version of it is not about hiding AI use at all. It is about proportionality. The radio host Erick Erickson said he had moved from Grammarly to Claude for proofreading, and objected that work he had written himself would now carry a mark saying Claude did it.

That case is genuinely hard to dismiss. A mark that cannot distinguish between drafting a piece and fixing its commas records the involvement of a tool, not the authorship of an idea. The counter-argument, also well represented in those threads, is that disclosure of AI involvement is the entire point and that proportionality is a problem for the detector to solve, not the marker.

Both can be true. A mark can be reasonable policy and still be blunt in individual cases.

04What this site does about it

Images

Re-encoded in your browser, which changes the format and takes the C2PA manifest, EXIF and XMP with it — the mechanism Anthropic named.

Removed

Text watermark

Untouched. It lives in word choice, and this tool does not rewrite your words. Deleting invisible characters has no effect on it whatsoever.

Not removed

What the text cleaner does remove is real but different: invisible Unicode characters, including the tag block that can hide an entire readable message inside a sentence. Useful, and not the same thing as the watermark.

Questions

Does Claude watermark text?

+

Yes. Since 11 August 2026 Anthropic applies an invisible watermark to Claude's text output across all models and products worldwide. Anthropic says the mark travels with the text when it is copied and pasted elsewhere. There is no opt-out.

Can the Claude text watermark be removed?

+

Not by deleting characters, because the mark is not a character. Anthropic states that text watermarks may be erased if the writing is heavily edited, which means substantial rewriting is the only known approach. Nobody outside Anthropic can verify removal in a given case, because the detection key is not public.

Does copying and pasting remove the Claude watermark?

+

No. Anthropic says explicitly that the mark travels with the text through copy and paste. Moving text between applications, documents or file formats does not affect it, because the signal is carried by the sequence of words rather than by any formatting.

Can the image metadata be removed?

+

Yes, and Anthropic acknowledges it. The company notes that a file's metadata could disappear if the format of the file changes. Re-encoding an image destroys the signed C2PA manifest along with EXIF and XMP, because the signature covers bytes that no longer exist.

Why is Anthropic doing this?

+

Transparency obligations under the EU AI Act are the stated driver, and Anthropic is applying the policy worldwide rather than only in Europe. Roughly two hundred companies had signed up to the relevant EU commitments by late July 2026.

Does this apply to Claude Code and the API?

+

Anthropic describes the policy as covering all Claude models and products worldwide, which reporting has taken to include coding and agentic products as well as consumer apps and API access.

Sources