You can inspect and remove selected hidden Unicode exactly. You cannot responsibly claim that this alone removes Claude’s statistical watermark.
1. Inspect before changing anything
Paste the text into the local inspector. It checks zero-width characters, byte order marks, unusual spaces, bidirectional controls, non-printing control characters, line separators, and variation selectors. These characters are not automatically malicious and are not proof of AI authorship.
2. Remove only known artifacts
Select a character type only after reviewing its name and locations. Zero-width joiners and variation selectors can affect emoji or script shaping; directional controls can change display order; unusual spaces may affect wrapping. Selective Cleaner removes exact selected code points and performs no Unicode normalization or silent replacement.
3. Verify the diff
The “Before” panel replaces removed invisible characters with readable markers. The “After” panel is the actual cleaned output. Review both, copy the result, then re-analyze it if needed.
4. Treat rewriting as meaning-changing
Anthropic’s official guidance says heavily edited, paraphrased, translated, or mixed text may no longer carry a detectable mark. This is a limitation of detection. It does not mean a particular rewrite guarantees removal, and rewriting can alter facts, tone, citations, or code behavior.
5. Keep provenance and policy obligations
Use cleanup for legitimate formatting, publishing, accessibility, and code-review needs. Do not present the tool as a way to defeat disclosure requirements, authorship policies, or integrity checks. Keep any provenance or disclosure your context requires.
For the current first-party description, read How Claude marks AI-generated content. For interpretation, compare the detector boundary, the text cleanup workflow, and the Claude Code guidance.
Clear answers, narrow claims
Claude watermark removal FAQ
Can hidden Unicode be removed from Claude text?
Yes. Exact Unicode code points can be identified and selectively removed. Review their locations first because some invisible characters affect layout, emoji, multilingual text, or source code.
Is hidden Unicode the same as Claude’s statistical watermark?
No. Anthropic describes an imperceptible model-level watermark woven into generated text. Hidden Unicode is a separate, directly observable layer.
Will rewriting remove a Claude watermark?
Anthropic says heavily edited, paraphrased, translated, or mixed text may lack a detectable mark. That is a limitation of detection, not a guarantee that a specific rewrite removes the mark.
How can I verify that removal worked?
You can verify exact character removal with the before-and-after diff and a second local scan. You cannot currently verify removal of Anthropic’s model-level mark here because the official detection mechanism is unpublished.
Inspect your content locally
See exact code points before choosing any cleanup.