AI Watermarks: Hidden Character Cleaner
Share tool
Only this public address is included. Your files, inputs and results stay out of the link.
Clean unwanted invisible characters, review the changes and copy the result.
Paste. Clean. Compare.
Processed on this deviceTip: Ctrl / ⌘ + Enter cleans your text. Regular Enter keeps a new line.
Selected: Remove invisible start markers · Remove optional hyphenation points
A visual example
Remove an artefact, keep the character
Some invisible characters are formatting leftovers. Others hold a written character or emoji together.
- Before cleaning
co[U+00AD]operate
👩💻
Make the difference visible
Here, [U+00AD] stands for an invisible soft hyphen inside a word. The brackets are only a visual explanation.
- After cleaning
cooperate
👩💻
1 soft hyphen removed
Keep the meaningful parts
With the default settings, the soft hyphen between these Latin letters is removed. The emoji's joiner stays intact.
An invisible character is not automatically an AI watermark.
Illustrative example. This example demonstrates character cleaning. It does not detect or verify the removal of statistical text watermarks. Protected code and other meaningful character sequences are treated separately.
How to use it
- Paste your text or try the example. The original stays beside the result.
- Use the defaults or choose specific changes in Cleanup settings.
- Clean the text and review the changes and any protected passages.
- Copy the current result or download it as a UTF-8 TXT file.
Compare the original and result before copying, especially if you enable optional space or zero-width character changes.
Details & limits
Up to 100,000 text units; emoji may count as several. Oversized or incomplete Unicode input is rejected without truncation. Processing stays in this browser. No text rewriting, watermark detection or original-file inspection.
What happens to code, links and quotations?
Recognized code, structured data, links and quoted spans are protected. Recognition uses cautious rules, not a complete parser. Keep the original and use a format-specific editor when exact syntax matters.
Can you remove AI watermarks from text?
This tool can remove selected invisible characters. It cannot detect statistical AI watermarks or verify their removal. Finding a hidden character does not establish who wrote a text; finding none does not establish that a text is watermark-free.
What a character change looks like
Illustrative example: [U+00AD] stands for one invisible soft hyphen. The bracketed label is shown here to make the difference visible; it is not literal input to remove.
- Before · character made visible
Please co[U+00AD]operate with the team.- After · cleaned text
Please cooperate with the team.
Default cleanup removes the soft hyphen between Latin letters in unprotected prose. The words are not rewritten.
Four things to distinguish
- Invisible characters
- Soft hyphens and special spaces can serve ordinary formatting. This cleaner shows their code points and applies only the changes you select. Unicode: Chapter 23
- Statistical text watermarks
- Patterns in token selection, such as SynthID, use a different mechanism. Removing character artifacts does not verify that such a pattern is gone. Statistical watermarks are not tested. Google DeepMind: SynthID
- File metadata and Content Credentials
- Information attached to a file is separate from its pasted text. This tool cannot inspect or remove metadata from your original document.
- Visible credits and source notices
- Bylines, source links and AI-use notices are visible content. Character cleanup does not remove them or rewrite their meaning. Editing does not by itself remove applicable attribution or disclosure obligations.
Looking for an AI watermark remover?
For text from any provider, this tool offers character cleanup only. It cannot verify watermark removal.
- ChatGPT
- OpenAI has researched text watermarking. Its current public verification offering covers images and audio, not pasted text. No supported ChatGPT text-watermark check is connected here; character findings cannot establish that an OpenAI watermark was present or removed. OpenAI: Content provenance
- Gemini
- Google describes SynthID watermarking in Gemini text through adjusted token probabilities. The pattern is in word choices, not extra invisible characters. This cleaner does not detect SynthID or verify its removal. Google DeepMind: SynthID
- Claude
- Anthropic documents SynthID-Text with its own key for supported Claude models, not hidden characters. Coverage depends on the model and rollout. Removing Unicode characters cannot verify removal of that signal. This tool has no access to Claude’s private-preview text detector. Anthropic: How Claude’s text watermark works
Two useful copy-and-paste fixes
A copied word will not match
A word copied from a document may contain a soft hyphen even when it looks normal. Inspect U+00AD, clean the text and retry the search or comparison.
A form needs ordinary spaces
Copied text may contain U+00A0 or U+202F instead of ordinary spaces. Replacement is optional and off by default: it changes line-breaking behavior, so review numbers, currencies and typography before enabling it.
AI text cleaner questions
What would a before-and-after watermark check show?
A suitable detector must check both texts using the same documented method and settings. Even a change from detected to not detected would mean only that this detector no longer recognizes its signal, not that all watermarks are gone. No suitable text detector is connected here, so this check is not performed. Anthropic: How Claude marks AI-generated content
Would rewriting remove a statistical watermark?
Substantial rewriting can weaken a statistical signal; small edits can leave it detectable. This does not guarantee removal or preserve meaning automatically. Another model may add its own watermark. No rewriting or provider-detector service is connected here, so this tool makes no before-and-after statistical claim. Google: SynthID Text method and limitations
Why not use the open-source SynthID detector?
The reference code needs the matching generation keys, tokenizer and configuration, plus suitable detector calibration. A detector configured for your own model does not test production Gemini or Claude watermarks. Those provider settings are not available to this tool; a lab example would not prove removal from your pasted text. Google: SynthID Text method and limitations
Can AI detectors still flag cleaned text?
Yes. Character cleanup is not a test of authorship or a way to certify human-written text. This tool runs no AI detector, provides no detection score and makes no promise about another service’s verdict. No rewriting service is connected here.
Which invisible characters can I remove?
Defaults remove a leading U+FEFF marker and soft hyphens (U+00AD) between Latin letters in unprotected prose. Optional changes cover U+200B, U+2060, interior U+FEFF and ordinary-space replacements for U+00A0, U+202F and supported typographic spaces (U+2000–U+200A, U+205F, U+3000). Joiners, non-joiners, direction controls, accents and emoji components stay intact.
Is my pasted text sent to an AI service?
No. Cleanup runs in this browser. Input and results are not submitted to a model, saved to browser storage or added to a URL. Copying uses your system clipboard; downloading creates a local TXT file. Device clipboard synchronization and backups are separate from this tool.