Disclaimer

Last updated 13 August 2026

No tool here removes Claude's text watermark

Since August 2026 Anthropic embeds a machine-readable watermark in Claude-generated text. It is applied during sampling inside the model, as a statistical pattern in word choice rather than as hidden characters. There is nothing inserted, so there is nothing to delete. No character cleaner on this site or anywhere else removes it.

Our rewriting tool changes word choice, which is the only mechanism that affects such a signal. We make no guarantee that it defeats detection. The detection threshold and key are not public, so nobody outside Anthropic can honestly promise a result — treat it as disruption, never as an erase.

Accuracy and currency

Our guides describe AI watermarking as we understand it at the date above, from published documentation and reporting. This area is changing quickly. Providers alter their systems, and details of how a watermark works are frequently undisclosed. Information here may become outdated or turn out to be incomplete. Verify anything you intend to rely on against the provider's own documentation, which we link to throughout.

Detection results prove less than people assume

Anthropic states that a detected mark indicates content may have been processed by Claude — which includes proofreading, translating or summarising something a human wrote. Absence of a mark equally proves nothing. AI-detection classifiers are a separate technology again, and have well-documented false-positive rates, particularly against non-native English writers. Nothing on this site should be used to accuse anyone of anything, or to reassure yourself that an accusation cannot be made.

Your obligations are yours

Academic institutions, employers, publishers and platforms set their own rules about disclosing AI assistance. Those rules apply to you regardless of what any tool does to your text, and removing a technical marker does not remove an obligation to disclose. Using these tools to breach such a rule is your responsibility, not ours. If you are unsure what applies, ask the institution before acting.

Not legal advice

We reference the EU AI Act, Article 50 and related transparency obligations to explain why watermarking exists. This is general information, not legal advice, and we are not lawyers. Obligations under the AI Act fall principally on providers and deployers of AI systems, and how they apply to you depends on facts we do not know. Take professional advice for your own situation.

Tool limitations

Mechanical synonym substitution has no understanding of your sentences and can produce awkward or wrong phrasing. Always read the changes before using the output.
Removing zero-width joiners splits multi-part emoji into their components. This is unavoidable — the joiner is what holds them together.
Re-encoding an image as JPEG or WebP re-compresses it and discards transparency. Use PNG when fidelity matters.
Metadata detection is a byte scan and reports what it recognises. It is not a cryptographic verification of a C2PA manifest.
Keep your own backups. We hold no copy of anything you process.

External links

We link to third-party sites for source material. We do not control them and are not responsible for their content or accuracy.