Claude AI Watermark Remover Tool
by KnowIt Ai
AI watermark remover
Find and strip invisible Unicode carriers from text and provenance metadata (C2PA, EXIF, XMP, generator tags) from images. Nothing leaves your browser.
Each image is inspected, cleaned, then inspected again so you can see what was removed and confirm nothing AI-related is left.
Updated September 2026 to cover Anthropic's Claude text watermark. If that announcement brought you here, jump to what it is and what this tool can and can't do about it.
What this tool does
Paste text from Claude, ChatGPT, Gemini, or any other AI tool, or drop in an AI-generated image, audio, or video file. The tool shows you every hidden carrier it finds, strips it out, and inspects the result again so you can confirm it's clean. Nothing you paste or upload leaves your browser.
Text
Finds and removes zero-width characters, word joiners, soft hyphens, bidi controls, tag characters, variation selectors, and space lookalikes. Inspect first, clean, then copy or save as .txt. Copy the full character report if you want the receipts.
Files
Strips C2PA Content Credentials, EXIF, XMP, and generator tags from AI-generated images, audio, and video. Works at the byte level, so pixels and audio are never re-encoded and there's no quality loss.
AI cadence
A quick heuristic score of how machine-like the prose reads: sentence-length burstiness, vocabulary variety, and stock AI phrasing. Useful for deciding whether something needs a rewrite. It is not a detector.
What the Text tab detects and removes
Every character below is either invisible or looks like an ordinary space but isn't. They survive copy and paste through almost any editor, which is why they're the usual place to hide a marker and why they break CSV imports, string matching, and code diffs.
| Character | Code point(s) | What it is | What happens |
|---|---|---|---|
| Zero-width space | U+200B | Invisible break point | Removed |
| Word joiner, zero-width no-break space (BOM) | U+2060, U+FEFF | Invisible no-break points; the BOM leaks in from file encodings | Removed |
| Zero-width non-joiner, zero-width joiner | U+200C, U+200D | Control how letters join | Removed, except where they hold emoji or Persian and Indic text together (option: strip all) |
| Soft hyphen | U+00AD | Invisible hyphenation hint | Removed |
| Invisible math operators | U+2061 to U+2065 | Function application, invisible times, separator, plus | Removed |
| Bidi marks and controls | U+200E, U+200F, U+202A to U+202E, U+2066 to U+2069 | Text direction control, also used for spoofing | Stray ones removed; right-to-left marks and paired embeddings kept by default (option: strip all) |
| Variation selectors | U+FE00 to U+FE0F, U+E0100 to U+E01EF | Glyph variant selectors | Removed, except emoji presentation selectors |
| Tag characters | U+E0000 to U+E007F | Invisible tag block used to smuggle hidden text | Removed, except inside flag emoji sequences |
| Blank-rendering fillers | U+3164, U+FFA0, U+180F | Hangul fillers, Mongolian free variation selector | Removed, kept where their script needs them |
| Reserved, noncharacter, and private-use code points | U+FFF0 to U+FFF8, U+FDD0 to U+FDEF, and others | Code points with no assigned meaning | Removed |
| Space lookalikes | U+00A0, U+2000 to U+200A, U+202F, U+205F, U+3000 | No-break, narrow no-break, em, en, thin, hair, and ideographic spaces | Normalized to a regular space (on by default) |
Options. Defaults preserve every invisible character that's doing real work: emoji joiners and presentation selectors, Persian and Indic ZWNJ, flag tag sequences, Hangul fillers, Mongolian FVS, Khmer inherent vowels, Arabic format marks, and right-to-left marks. Two options strip those too. NFKC normalize folds ligatures, superscripts, and fullwidth forms into plain characters. Replace Latin confusables turns Cyrillic and fullwidth lookalikes into ASCII, which changes visible text, so it's off by default.
How to remove hidden characters from AI text
- Paste your text into the box above, or drop a .txt, .md, .html, .csv, or .json file on it. It can come from claude.ai, ChatGPT, Gemini, the API, Claude Code, or a document someone sent you.
- Click Inspect only to see what's in there before anything changes. Every carrier is listed with its code point, name, and count.
- Click Clean, then Copy or Save .txt. Copy report gives you the before-and-after list if you need to show someone what was removed.
The visible words don't change. Try sample loads a piece of text seeded with hidden characters so you can watch it work.
How to strip C2PA and metadata from AI images and files
- Open the Files tab and choose or drop your files. Images: PNG, JPEG, WebP, AVIF, HEIC, BMP, GIF, TIFF. Audio and video: MP4, MOV, WAV, MP3, FLAC. Several at once is fine.
- Each file is inspected, cleaned, and inspected again. You see exactly which containers held C2PA, EXIF, XMP, or generator tags and confirm nothing AI-related is left.
- Tick Keep non-AI metadata if you want camera EXIF to stay, then Download all.
The cleaning is byte-level: PNG chunks, JPEG APP segments, WebP RIFF chunks, AVIF, HEIC, MP4, and MOV boxes, WAV chunks, MP3 ID3v2 frames, the FLAC C2PA carrier, GIF extension blocks, TIFF tags, and BMP trailing bytes. Pixels and samples are never re-encoded. Color profiles stay either way.
Does Claude watermark its text?
Yes. Since August 2026, Anthropic embeds a watermark in text generated by Claude models released on or after August 2, 2026, and it's working to add the watermark to older models as well. Anthropic made the change to comply with Article 50 of the EU AI Act, which took effect on August 2, 2026, but the watermark applies worldwide and there's no opt-out. It covers claude.ai, the Claude API, Claude Code, Claude Cowork, and Claude on AWS, Google Cloud, and Microsoft Foundry.
What the watermark is. Anthropic describes it as woven into the text through the words the model chooses. It adds no extra characters and no extra tokens, and it doesn't change the meaning or readability of the output. It's designed to survive copy and paste and light editing. Anthropic also says where it's weak: short passages, factual text where only one word fits, and code, where it mostly lives in comments. A full rewrite removes it. Translations are fully watermarked because the model picks every word.
Files are different. Images and other files Claude generates get digitally signed provenance metadata using the C2PA standard. Anthropic's own documentation notes that this metadata can be stripped by re-saving, converting, or screenshotting the file. The Files tab above removes it without re-encoding anything.
Who can detect it. Anthropic's detection API is in private preview for regulators, law enforcement, media, fact-checkers, researchers, educational organizations, EU civil society groups, and enterprises with their own compliance obligations. There's no public checker yet. Anthropic also says the watermark identifies Claude, not you: it carries nothing about the user, their organization, or their chats, and it says nothing about who owns the text.
What this means for this tool. "Claude watermark" now means two different things:
- Character-level carriers. Zero-width characters, word joiners, soft hyphens, unusual spaces, tag characters, and the C2PA metadata on files. These are real, they're detectable, and this tool finds and removes all of them and shows you the result is clean.
- Anthropic's word-choice watermark. No character-level tool can remove this, because there's nothing to delete. The only thing that changes it is choosing the words again. If you need that layer gone, clean the text here first, then rewrite it with a different model.
Does ChatGPT, Gemini, or Grok watermark text?
Status as of September 2026:
| AI tool | Text watermark | Details |
|---|---|---|
| Claude (Anthropic) | Yes | Models released from August 2, 2026; older models being added ahead of the EU's December 2, 2026 deadline. Word-choice watermark plus C2PA metadata on files. Global, no opt-out. |
| ChatGPT (OpenAI) | Not confirmed | OpenAI signed the EU transparency code and says it plans to extend provenance signals to text, but hasn't announced a deployed text watermark. Images from ChatGPT carry C2PA metadata and a SynthID mark. |
| Gemini (Google) | Yes | SynthID text watermark since 2024. Statistical, word-choice based. Gemini images also carry a SynthID pixel watermark, which this tool does not remove. |
| Grok (xAI) | No | xAI did not sign the EU transparency code. |
None of these text watermarks are Unicode-based. That's why "no hidden characters found" doesn't mean "no watermark," and why the hidden characters in AI text usually come from somewhere else.
About the AI cadence score
The AI cadence tab gives prose a score based on three things: how much sentence length varies (human writing is bursty, model writing is even), how varied the vocabulary is, and how much stock AI phrasing shows up. It's a heuristic for deciding whether a passage needs a rewrite. It is not an AI detector, it says nothing about watermarks, and it can be wrong in both directions.
Nothing leaves your browser
Text and files are processed with JavaScript on your device. Nothing is uploaded, logged, or stored, and there's no account to create. Close the tab and it's gone. The engines are MIT-licensed ports of the open-source watermarks-remover project, parity-tested against the original Python service. Use the tool on content you own or are authorized to modify.
Frequently asked questions
What is an AI text watermark?
An AI text watermark is a signal built into AI-generated text so a detector can later tell that a model produced it. There are two kinds. Character-level marks use invisible Unicode characters such as zero-width spaces. Model-level marks, like Google's SynthID and Anthropic's Claude watermark, are statistical patterns in word choice and add no characters at all.
Does Claude watermark its text?
Yes. Anthropic embeds a watermark in text generated by Claude models released on or after August 2, 2026, and is adding it to older models. It applies worldwide, has no opt-out, and lives in word choice rather than hidden characters. Anthropic's detector is in private preview for regulators, researchers, media, and compliance teams.
Can this tool remove Anthropic's Claude watermark?
It removes every character-level carrier in the text and the C2PA metadata on files: zero-width characters, hidden Unicode, unusual spaces, tag characters, and provenance manifests. It cannot remove a word-choice watermark, because there is no character to delete. Anthropic says only rewriting the text changes that layer. Clean here first, then rewrite with a different model if you need to.
Does ChatGPT watermark its text?
Not that OpenAI has confirmed. OpenAI signed the EU transparency code in 2026 and says it plans to extend provenance signals to text, but as of September 2026 it has not announced a deployed text watermark. Images from ChatGPT do carry C2PA metadata and a SynthID mark. Hidden characters in pasted ChatGPT text usually come from the chat interface, not a watermark.
Does Gemini watermark its text?
Yes. Google has watermarked Gemini text with SynthID since 2024. Like Claude's watermark, it is statistical and lives in word choice, so a character cleaner will not detect or remove it.
What are zero-width characters?
Zero-width characters are Unicode characters that take up no visible space. The common ones are the zero-width space (U+200B), zero-width non-joiner (U+200C), zero-width joiner (U+200D), and word joiner (U+2060). They have legitimate uses in some scripts and in emoji, but they also survive copy and paste, which makes them a popular place to hide a marker.
How do I check text for hidden characters without changing it?
Paste your text and click Inspect only. You get a report of every carrier found, with its Unicode code point, name, and count, and nothing is modified. Copy report puts that list on your clipboard.
Will it break emoji or non-English text?
Not with the default settings. Some invisible characters are load-bearing: zero-width joiners hold multi-part emoji together, and zero-width non-joiners are required in Persian and some Indic scripts. The defaults keep those in place and only remove characters that carry no meaning where they sit. The two strip options in the Options panel remove them too, so leave those off unless you know you want that.
What does the Files tab remove from images?
C2PA Content Credentials manifests, EXIF, XMP, and generator tags from PNG, JPEG, WebP, AVIF, HEIC, BMP, GIF, and TIFF images, plus the same carriers in MP4, MOV, WAV, MP3, and FLAC files. It edits the file containers directly, so pixels and audio samples are never re-encoded and there is no quality loss. Color profiles are kept, and camera EXIF can be kept with the Keep non-AI metadata option.
Does it remove SynthID or other pixel watermarks from images?
No. SynthID for images, StegaStamp, and Tree-Ring live in the pixels themselves, not in metadata. This tool only removes metadata and leaves pixel data untouched.
Is the AI cadence score an AI detector?
No. It is a heuristic that combines sentence-length burstiness, vocabulary variety, and the density of stock AI phrasing. It helps you judge whether prose reads as machine-written and needs a rewrite. It says nothing about watermarks and can be wrong in both directions.
Is it legal to remove AI watermarks?
Generally yes for content you created or have the right to use. Removing invisible characters from your own document is ordinary text cleanup, and Anthropic's own documentation says its watermark says nothing about ownership or authorship. Follow the terms of service of the AI tool you used, disclose AI use where a school, employer, publisher, or law requires it, and do not use cleaned content to misrepresent someone else's work. This is not legal advice.
Which AI tools does this work with?
All of them. The tool looks at characters and file containers, not at which model produced the content, so it works on output from Claude, ChatGPT, Gemini, Copilot, Grok, Perplexity, DeepSeek, Llama, and anything else. It also cleans text copied from Word, Google Docs, PDFs, and web pages, which pick up the same hidden characters.
Is it free, and does my content get uploaded?
It is free, there is no account, and nothing is uploaded. Text and files are processed in your browser and never sent to a server.
Free Claude watermark remover and checker by KnowIt Ai. Find and strip zero-width characters, hidden Unicode, and C2PA metadata from Claude, ChatGPT, and Gemini text and files.