Claude watermarks
Claude watermarks and content credentials, honestly
Marks are evidence, not verdicts. Editing is not removal. The twelve pages below cover what Claude does and what it cannot do.
Claude now marks its output. Models launched in the EU on August 2, 2026 carry text watermarks applied at the model level. Generated files like .svg, .png, and .jpg carry signed C2PA provenance metadata. The marks are real. There is no off switch, and the rollout is ongoing. Anthropic also documents the limits. Marks can be lost to heavy editing, paraphrasing, translation, screenshots, and format conversion. A detected mark is not conclusive. The absence of a mark does not mean content was human. The honest response is to edit the text until it is yours, and to disclose where you must.
What Claude marks
Claude models launched in the EU from August 2, 2026 carry machine-readable marks at launch. Text gets an imperceptible watermark at the model level, and it travels with copy and paste. Generated files like .svg, .png, and .jpg get signed C2PA provenance metadata.
Where the marks live
Marking covers Claude products and major cloud platforms, worldwide. Some platforms or features may not support every mark type. Plain text has no C2PA container, so file provenance only exists in formats that support it.
The limits Anthropic documents
Heavy editing, paraphrasing, or translating can degrade a text mark. So can very short passages. Format conversion, re-saving, or screenshots can strip metadata. A detected mark isn't proof on its own, and its absence doesn't prove a text is human. The detection tooling isn't out yet.
No off switch
As of August 2026, Anthropic's documents list no setting, no admin control, and no opt-out for marking. Older models are being updated as part of the rollout.
The legal frame
Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content. The EU AI Act's transparency obligations start on August 2, 2026. Systems placed on the market before that date get until December 2, 2026.
The honest response
Edit the draft until it reads as yours. Keep the facts you supplied. Where a school, employer, or law requires disclosure, disclose. Anthropic's usage policy treats presenting AI work as your own as a breach.
The twelve pages
- Does Claude watermark text? Yes, and here is what that means
- Claude text watermarks are real and rolling out. What Anthropic applies, where it applies, what survives editing, and why a detected mark is not proof.
- Can you remove a Claude watermark? The honest answer
- There is no switch that deletes a Claude text watermark, and we won't sell you one. What the marks actually are, and what honest editing does instead.
- Claude content credentials: what they are, and are not
- On Claude, content credentials means C2PA provenance metadata on generated files, not leaked API keys. What Claude attaches, to which files, and how to check.
- C2PA content credentials, explained simply
- What C2PA is, what Content Credentials contain, why they are not DRM, and why the first independent audit said the standard should not yet be relied on.
- Stripping C2PA metadata: what it does, and does not
- C2PA metadata can be dropped from most files, and it is honest to say so. What stripping achieves, what it cannot achieve, and why that is the whole story.
- Claude and AI detectors: what they actually flag
- Detectors do not read watermarks and cannot name the model. They score predictability, with bias and false positives. What that means for Claude text.
- Why Claude text gets flagged as AI: not always a watermark
- A flag is usually a statistical judgment about style, not a watermark match. The patterns detectors key on, and the honest human writing they misread.
- How to rewrite Claude text in your own voice
- A concrete edit pass for Claude drafts: strike the model's reflexes, keep your facts, shorten the sentences, and make the claims yours. No tools required.
- Make Claude writing sound human: edit, do not disguise
- The human sound comes from editing, not prompting for a disguise. Which patterns to remove, which prompts help, and how to keep your own voice in the draft.
- Claude's writing style: the patterns that give it away
- Claude's default style has recognizable patterns: validation openers, list scaffolding, dense dash chains. How to spot them, strike them, and keep your voice.
- How AI writing detection works, and why it is not proof
- Detectors score statistical predictability, not authorship. How they work, their accuracy record, their documented bias, and why a score is never proof.
- EU AI Act text watermarking: what writers need to know
- The EU AI Act's transparency obligations apply from August 2, 2026, and the watermarking code is signed. What that means for writers, and one delay to know.
Questions
Does Claude watermark text?
Yes. Claude models launched in the EU from August 2, 2026 carry imperceptible text watermarks applied at the model level. Older models are being updated.
Can Claude watermarks be removed?
Anthropic lists the ways marks get lost. Heavy editing, paraphrasing, translation, and very short passages weaken them. Format conversion, re-saving, or screenshots strip the file metadata.
What does C2PA mean on Claude files?
Generated .svg, .png, and .jpg files carry signed provenance metadata that follows the C2PA open standard. It signals that a file was processed by Claude and lets you check whether it was tampered with.
Is there an off switch for the watermark?
No. Anthropic's documentation lists no setting, no admin control, and no opt-out for marking. Older models are being updated as part of the rollout.
underslop is for editing, formatting, grammar, and privacy. It is not designed to hide AI-generated content. We make no promise about what any detector, watermark, or provenance check will report. You are responsible for how you use it, and for following your school's, employer's, or instructor's rules. Do not use it to misrepresent authorship. Presenting AI-assisted work as your own without permission or attribution can breach your institution's integrity policy. It can also breach the usage policy of the tools you used. AI-generated files may carry provenance metadata or watermarks, and format conversion, re-saving, or editing can strip file metadata. AI providers acknowledge that limitation. We do not remove, alter, or forge provenance metadata or watermarks as a service. We are not affiliated with Anthropic, any AI provider, or any detector vendor. Detector scores are estimates, not proof of authorship. We are not liable for how you use the output, including academic disciplinary actions. If your school allows AI assistance with disclosure, use it transparently. If it does not, do not use it to misrepresent your work.