Free Claude Watermark Remover: Clean and Rewrite Your Draft

Two different things, and they come out in two different ways. Read this and the next two answers before you trust any tool that calls itself a watermark remover, including this one.

Hidden characters are real Unicode codepoints sitting inside your text. They are listable, and the free tool above takes them out completely. Claude's watermark is a pattern in which words the model chose, so there is no character to delete: it comes out by replacing the words, which is what a full rewrite does.

Formatting artifacts it can point to in your text, split into two groups, because some invisible characters are junk and others are doing a job. Low-risk ones are cleaned automatically. Joiners, direction controls, variation selectors and soft hyphens are listed for you to approve one by one, because they carry meaning in many languages.

• Zero-width characters · zero-width space, zero-width joiner and non-joiner, word joiner, byte order mark
• Unicode tag characters · a block that can encode a full hidden ASCII message inside ordinary-looking text
• Bidirectional controls · left-to-right and right-to-left marks, embeddings, and overrides
• Soft hyphens and variation selectors · leftovers that survive copying between apps
• Exotic whitespace · non-breaking, narrow no-break, hair, thin and figure spaces, normalized to a plain space

These turn up in text from every AI tool, not just Claude, and also from Google Docs, Word, and web pages. They can break word counts, corrupt search, and look suspicious in a plagiarism system that renders raw characters.

The automatic pass never rewrites a word: it removes invisible characters, fixes spacing, and, unless you turn the toggle off, normalizes smart quotes and dashes. Anything that could change how a word reads is held back for you to approve. Characters that are provably doing their job, like emoji sequences, flag emoji, or a joiner between two Persian letters, are never flagged at all.

Yes, by rewriting it. Not by cleaning characters, which is all the free tool above does: deleting characters cannot touch a pattern that was never stored in characters in the first place.

Anthropic's own published position is that light editing probably will not remove the mark, while a complete rewrite that replaces every word will. That is exactly what Lunchbreak's rewrite does, at the sentence level rather than swapping synonyms, and our rewrite provider has validated that it removes the Claude watermark.

Anthropic began watermarking Claude's text output in August 2026 to satisfy the EU AI Act's transparency rules. Anthropic has since published how it works: a key, plus the few words before each choice, settles which word the model picks where several would do equally well. It is a pattern in word choice, not a character sitting in your text, which is why deleting characters cannot touch it.

Anthropic has also said a watermark detection API is coming. When it lands, anyone will be able to check a draft themselves. Watermarking is new and moving quickly, so we will update this page as the picture changes.

Hidden characters, which this tool removes:

• Real Unicode codepoints stored inside your text
• Listable by any program that inspects raw characters
• Removable completely, and safely, right here
• Arrive from any source: Claude, ChatGPT, Google Docs, the web

Claude's watermark, which takes a rewrite:

• An undocumented signal applied while the model generates the text
• Not visible, and not listable by any tool available to you
• Removed by replacing the words, which is what a full rewrite does, not by deleting characters
• Readable only by Anthropic, and whoever it grants access to

So the two steps do different jobs, and you want both. Cleaning fixes the formatting debris sitting in your text. Rewriting replaces the word choices the mark lives in, which is what removes it.

Anthropic has now published how it works, so this is no longer guesswork. A key, combined with the few words preceding each choice, decides which word the model picks wherever several would work equally well. The signal lives in those choices, so it survives copying and pasting and cannot be found by inspecting characters.

What has been reported:

• It applies to Claude models released after 2 August 2026, across the API, Claude, Claude Code, Claude Cowork, and Claude Tag
• It is invisible and Anthropic says it does not change how the writing reads
• It is designed to survive copying and pasting
• Coverage says it “may persist through some editing,” which is not a number

• Light editing probably will not remove it, while a complete rewrite that replaces every word will

What is still moving:

• A watermark detection API is coming, though it is not available yet, and it is unclear who will get access
• Watermarking is being extended to models released before 2 August 2026 over the coming months

Watermarking is new and changing quickly, so we will update this page as the picture changes. Last checked: 18 August 2026.

The full breakdown of Claude's watermark →

It helps with real problems. It does not help with the one people assume.

It does help if:

• Your word count or character count is off for no obvious reason
• Text is breaking layouts, search, or form validation
• You are pasting between Google Docs, Word, Notion, and a submission portal
• You want text that is clean at the byte level before you submit it

It does not help if:

• You are trying to remove Claude's model-level watermark, which takes a rewrite, not a cleaner
• You are trying to pass an AI detector

That second one matters, because it is the reason most people land on a page like this. AI detectors do not read watermarks or hidden characters. They analyse writing patterns: sentence rhythm, word choice, how predictable each next word is. A document with zero invisible characters can still be flagged, and human writing gets flagged all the time.

If the watermark or a detector is the actual worry, cleaning characters is the wrong lever. Rewriting is the one that moves.

The step you can genuinely control is knowing how your writing reads before someone else runs it through something.

• AI checker · run your writing against major AI detectors at once and see each result separately. Detectors disagree with each other more than people expect, and the spread tells you more than a single number. Free to check.
• AI humanizer · rewrites for tone, rhythm, and flow so writing reads like a person instead of a template. You compare the original against the rewrite and keep what you want.

On watermarks specifically: a complete rewrite that replaces every word is what removes the mark, which is what the humanizer does. See what an AI watermark is and what it proves.

Yes, every single change. Each artifact is marked in your own text and itemised by category with its Unicode codepoint and position. Anything that could change how a word reads, like a joiner or a soft hyphen, is held back for you to approve one at a time.

No. The scan runs locally in your browser and no account is needed to use it. Your draft only reaches us if you choose to save it into your Lunchbreak account, and the note under that button says so at the point you decide.

438, counted from the scanner's own tables rather than quoted from marketing. That covers zero-width characters, the Unicode tag block, bidirectional controls, variation selectors, private-use characters and sixteen different space characters.

Mostly in what it refuses to claim. It removes formatting artifacts and itemises every one; a typical watermark remover claims to remove Claude's watermark and shows you nothing. It runs in your browser rather than on someone's server, it is free with no account, and it tells you plainly that cleaning characters is not what removes the watermark. Rewriting is, and that is the step it hands you on to.

No. Joiners that are doing real work in Persian, Arabic and Indic scripts are left alone, as are emoji sequences, flag emoji, and Chinese, Japanese and Mongolian variation selectors. Cleaning your text should never break someone's name.

Yes. No account, no card, no limit. It cleans hidden formatting characters, not Claude's watermark.

Yes. Anthropic began embedding a watermark in Claude's text output in August 2026, applied at the model level across its products.

For hidden characters, yes. It strips invisible codepoints from any text regardless of which tool produced it. OpenAI had not announced text watermarking equivalent to Anthropic's at the time of writing.

Same answer. The cleaner is model-agnostic because invisible characters are just Unicode. Google had not announced comparable text watermarking when this page was last checked.

No. Detectors analyse writing patterns, not hidden characters. Cleaning your text changes nothing about how a detector scores it.

Mostly no. Plain text editors preserve Unicode codepoints, including the invisible ones. That trick strips formatting, not characters.

Almost always. Everything removed automatically is invisible, a whitespace variant replaced by a normal space, or a private-use character, which has no standard meaning and may show as a box or a custom icon. Every change is listed in the report, so you can see exactly what went.

A block of codepoints, U+E0000 to U+E007F, that mirrors ASCII but renders as nothing. A full hidden message can be encoded in them and pasted inside normal-looking text. This tool removes them, unless they form part of a real flag emoji.

• What an AI watermark is, and what it proves
• Does Claude watermark AI text?
• What is AI text watermarking?
• Can Turnitin or GPTZero detect watermarked text?
• Why AI detectors flag human writing
• What to do if you get flagged for AI