Explainer

How Claude watermarks work

As of 2 September 2026, Claude watermarks text from supported models launched on or after 2 August 2026 — currently Fable 5.1 and Mythos 5.1. The official mark is statistical, with no hidden characters. Detection is private preview. This page covers the text mark, C2PA files, and rewrite.

Longer essay: Claude watermarks explained. Launch story: August 2026 news.

Which models are marked

Anthropic currently names Fable 5.1 and Mythos 5.1 as supported. New Claude models launched on or after 2 August 2026 get machine-readable marking at launch. Models from before that date are not confirmed marked — Anthropic says a retrofit is still in progress.

Official details: How Claude marks AI-generated content.

What the text mark actually is

During generation, the model’s word choices can be biased toward a pattern a verifier can later score (a SynthID-Text–style statistical mark). Readers still see normal English; machines may see a Claude-shaped statistical trail. Anthropic stresses that nothing is added to the text and there are no hidden characters— the watermark is part of the wording, not sidecar metadata or invisible Unicode. Clearing styles or exporting to .txt does not remove it.

Because the mark is part of the wording, copying Claude output into Docs, Notion, email, or code comments can carry the signal with it. Light synonym swaps often keep enough of the original sentence skeletons for a detector to still score.

When the signal is weakest

Capacity depends on entropy. Highly constrained answers (“What is 2+2?”), short tweets, and dense code blocks leave fewer free choices for a sampler to bias. Long blog-style prose is the harder case — and the main reason people search for a Claude watermark remover.

Signed provenance metadata on files

For supported files such as PNG, JPG, and SVG, Claude can attach signed provenance metadata using the C2PA open standard. That label can signal a file was processed by Claude and help detect tampering.

C2PA is not the same as the invisible text mark. Re-encoding an image may drop metadata; pasting a paragraph into Notion will not “drop” a text watermark the same way. Strip PNG, JPEG, or SVG credentials in your browser: remove C2PA. Details: Claude C2PA provenance explained.

What survives which edits

Change you make	Text watermark	C2PA on files
Copy/paste the text	Usually keeps it	N/A
Light synonym edits	Often keeps it	N/A
Heavy paraphrase / rewrite	Strongly weakened	N/A
Clear formatting / export .txt	Does not remove it	N/A
Re-encode / strip metadata	N/A	Often drops it
Screenshot of an image	N/A	Usually drops signed metadata

That matrix is why a Claude watermark remover for prose is a rewrite tool, not a “delete hidden character” tool — and why file provenance needs a different workflow.

Why Anthropic shipped this (EU AI Act)

Transparency rules push providers to make AI-generated content identifiable in machine-readable ways. Anthropic’s public stack is text watermarks plus C2PA on supported files, worldwide on supported models — not a public “AI” badge on every paste.

Machine-readable marks in Claude-generated text
Signed provenance metadata on supported files
Detection in private preview for eligible organizations — not a public consumer detector

Using Claude does not by itself mean every downstream paste must show a public AI label. Provider-side marks and your disclosure duties (if any) are separate questions. Neutralizing brand-owned copy is different from hiding provenance where disclosure is required.

What detection means (and doesn’t)

Anthropic says a detected mark indicates content may have been processed by Claude — not full proof of authorship. Short, heavily edited, translated, or pre-marking model outputs may not carry a reliable signal.

Watermark detection is a private preview for eligible organizations (regulators, certain enterprises, and similar). There is no public consumer detector. Treat watermarking as a generation-time property of supported-model outputs, not something you can eyeball in the UI. You can inspect leftover traces (HTML class names, zero-width characters, exotic spaces). Those leftovers are not Anthropic’s statistical mark, and this site does not detect Fable 5.1 or Mythos 5.1 watermarks. A detect → rewrite → re-check loop is not available until a public API exists.

Common misconceptions

“It’s a hidden Unicode character.” No. Anthropic’s official text mark is statistical (SynthID-Text family). Nothing is added to the text and there are no hidden characters. Zero-width characters and artifact HTML class names are a separate leftover our free checker can inspect — not the official mark.
“Turning off formatting removes it.” Formatting is unrelated to sampling bias.
“C2PA tools fix text.” C2PA is for files; see Claude C2PA.
“GPTZero is Anthropic’s detector.” Classifiers score AI-like writing. They are not Claude’s watermark verifier. See GPTZero, Turnitin, and ChatGPT.

FAQ

Is a Claude AI text watermark visible?

No. Readers see normal prose. The signal is statistical over word choice, not a logo or hidden Unicode string you can select and delete.

Does clearing formatting remove a Claude text watermark?

No. Formatting is unrelated to sampling bias. Exporting to plain text still carries the words that encode the mark.

When is the watermark weakest?

Very short snippets, highly constrained factual answers, and code-heavy outputs often carry less reliable signal than long free-form essays.

Did the EU AI Act force Anthropic to watermark Claude?

Transparency rules pushed providers toward machine-readable marks. Anthropic shipped text watermarks plus C2PA on supported files. Your own disclosure duties, if any, are a separate legal question.

Which Claude models are watermarked?

Anthropic currently names Fable 5.1 and Mythos 5.1. New models launched on or after 2 August 2026 are marked at launch. Models from before that date are not confirmed marked; a retrofit is still in progress.

Check traces for free Remover vs paraphrase tools →

Related guides

Source framing based on Anthropic’s help article on how Claude marks AI-generated content (updated 2 September 2026). Not affiliated with Anthropic. Updated 2 September 2026.