Remove AI Watermarks From Text
Text watermarks live in word choice, so the only thing that removes them is a genuine rewrite. Here's what works, what doesn't, and where the lines are.
Try it now — it's free| Method | Removes text watermark? | Why |
|---|---|---|
| Deleting hidden / zero-width characters | No | Modern watermarks aren't characters. They're a statistical pattern in the words themselves. |
| Copy-pasting as plain text | No | The words stay the same, so the pattern stays too. Anthropic notes the watermark travels with copied text. |
| Light manual edits | Partially | Changing a few words weakens the signal, but long passages usually keep enough to be detected. |
| Synonym-swapping tools | Partially | Word-level swaps leave sentence structure intact and often read unnaturally. |
| Full rewrite in new wording | Yes | Replacing the word choices replaces the pattern. Anthropic says a complete rewrite will remove its watermark. |
Why most 'watermark removers' don't work on text
Many tools advertised as AI watermark removers only strip invisible Unicode characters such as zero-width spaces. That was never how Google, Anthropic or other major providers watermark text. Their watermarks (SynthID Text and Claude's, which is based on it) are encoded in which words the model chose. As long as those words remain, the signal remains, even after copy-pasting, reformatting or removing special characters. To remove it, the text has to be rewritten into genuinely different wording.
How to rewrite watermarked AI text
Paste your text
Drop in text from Claude, Gemini, ChatGPT or any other AI model.
Humanize it
Our humanizer rewrites the text at the sentence level with new structure and word choice, not just synonym swaps.
Make it yours
Add your own examples and voice, then check the result with the built-in AI detector.
Tips for a thorough rewrite
-
1
Rewrite, don't tweak
Research consistently shows light edits leave most of a watermark intact. Restructure sentences and change the vocabulary throughout.
-
2
Work in chunks
Process long documents in sections of a few hundred words so every part gets a full rewrite.
-
3
Add your own substance
Your own examples, opinions and details make the text more genuinely yours and more useful to readers.
-
4
Remember metadata for files
Images and documents may also carry C2PA Content Credentials. These are separate from the text watermark and are stripped by screenshots or re-exporting.
Use it responsibly
Rewriting AI-assisted drafts into your own voice is common and legitimate. But watermarks exist to support transparency, and some rules still apply after you rewrite. Schools and employers may require you to disclose AI use regardless of watermarks. AI providers' terms may prohibit tampering with their markings. The EU AI Act requires deployers to disclose AI-generated text published to inform the public on matters of public interest. China prohibits removing required AI labels. Rewriting text never makes it acceptable to pass off deepfakes or deceive people about AI-generated media. If in doubt, disclose.
Remove AI Watermark FAQ
Anthropic itself says light editing probably won't remove the watermark completely, but a complete rewrite where every word is replaced will. Research on the deployed system also found that paraphrasing or translating with a different model substantially weakens the signal.
It depends on where you are and what the content is. Rules in the EU, China and California mainly target AI providers and platforms, and require them to mark content and discourage tampering. China explicitly bans malicious removal of AI labels. Rewriting your own text is generally treated differently from stripping labels from media meant to deceive. Check your local rules and any school or workplace policies.
Watermarks and AI detectors are separate systems. A thorough rewrite addresses both, since detectors look at writing style (perplexity and burstiness) rather than a hidden key. Use the detector on this page to check your result.
Translating with a different, unwatermarked tool largely removes it, according to academic research. Translations produced by Claude itself are watermarked again, since Claude generated the new text.
As of September 2026, ChatGPT text is not known to carry a statistical watermark, though OpenAI has said it plans to add text provenance. Gemini and Claude text are watermarked.
Sources
- Anthropic: How Claude's text watermarking works
- Krishna et al., Paraphrasing evades detectors of AI-generated text (DIPPER, 2023)
- Sadasivan et al., Can AI-Generated Text be Reliably Detected? (2023)
- Zhang et al., Watermarks in the Sand (2023)
- Nemecek et al., evaluation of deployed text watermarks (2026)
- China Law Translate: AI content labeling measures