Anthropic began watermarking Claude text on 14 August 2026. GPTCleanup scans your draft for the AI-sounding patterns and giveaways that make it read as Claude, rewrites them so your writing sounds natural, and removes the hidden Unicode characters copied Claude text carries.
Cleaned drafts
end up here
Medium
Substack
WordPress
Ghost
Hashnode
dev.to
Quora
Reddit
WattpadPlatform names and marks indicate where drafts cleaned with GPTCleanup are commonly published. All marks are the trademarks of their respective owners. No partnership, sponsorship, or endorsement is implied.
The Claude watermark detector scans pasted text for the invisible Unicode markers AI detection systems look at: zero-width spaces, zero-width joiners, non-breaking spaces, and formatting attributes picked up from the Claude interface.
An AI watermark remover, hidden character remover, and invisible character cleaner in one pass. Word joiners, byte-order marks, soft hyphens, variation selectors, and tag characters all come out while your visible words stay exactly as written.
Every watermark character is reported by Unicode code point with a count, the way a proper Claude watermark checker should, and each flagged phrase is quoted back with the pattern it matched.
Em dashes, en dashes, two-em dashes, and three-em dashes are detected separately from hidden characters, alongside the long qualifying clauses and hedges Claude reaches for more than any other model.
As of 14 August 2026 it does, and this changed recently enough that most pages selling a Claude watermark remover have not caught up.
Anthropic announced text watermarking for Claude and began applying it to new models from 14 August 2026. It is switched on globally rather than only in the EU, and it is applied at the model level, so it reaches text from the API, the Claude apps, and Claude Code alike. Older models released before August are being covered over the following months.
It works by changing where the randomness in word choice comes from. When several words would do equally well, a key plus the preceding words decides which one the model picks, leaving a statistical pattern across the passage. Anthropic’s example is choosing between “overcast” and “grey”. It is the same family of approach as Google DeepMind’s SynthID-Text, going back to a 2022 proposal by Scott Aaronson.
What GPTCleanup does with Claude text is the part a reader, and a detector, actually notice. It scans your draft for the AI-sounding patterns and giveaways, rewrites them so your writing reads naturally, and removes the hidden Unicode characters that copied Claude text carries.
Anthropic is candid about the limits. Light editing, in their words, “probably won’t remove the watermark completely; a complete rewrite where every word is replaced will.” Detection also “doesn’t work well on small samples” and is sparser on factual passages, and it cannot tell text Claude wrote from text Claude heavily edited. A detection API is promised but not yet released.
Claude text carries two kinds of giveaway: the hidden characters that come along when you copy it, and the AI-sounding phrasing that makes it read as Claude. GPTCleanup handles both.
The first is the invisible Unicode that travels with copied Claude text: zero-width spaces, non-breaking spaces, word joiners, and formatting attributes picked up from the interface. Those are real, they break find-and-replace and screen readers, and GPTCleanup removes them.
The second thing this page finds is the phrasing. Claude hedges more than any other major model: it qualifies before it commits, stacks clauses inside one sentence, and reaches for paired em dashes to hold them together. None of that is hidden, which is why it survives character cleanup, and it is what a reader reacts to and a detector scores.
421 characters in total: 37 named individually below, plus 384 more across the variation selector and tag-character ranges. The set is shared across models, because these characters come from interfaces and editors rather than from Claude itself.
No width at all. They survive copy and paste, they are invisible in every editor, and they are what people usually mean by a hidden watermark character.
Removed completely
U+200BZero-width spaceU+200CZero-width non-joinerU+200DZero-width joinerU+2060Word joinerU+FEFFByte-order markU+00ADSoft hyphenU+180EMongolian vowel separatorSpaces that are not the ordinary space character. Identical on screen, but they break find-and-replace, wrap oddly, and split words in search indexes.
Replaced with a normal spaceU+00A0Non-breaking spaceU+202FNarrow no-break spaceU+205FMedium mathematical spaceU+3000Ideographic spaceU+2000En quadU+2001Em quadU+2002En spaceU+2003Em spaceU+2004Three-per-em spaceU+2005Four-per-em spaceU+2006Six-per-em spaceU+2007Figure spaceU+2008Punctuation spaceU+2009Thin spaceU+200AHair spaceU+2028Line separatorU+2029Paragraph separatorCharacters that look exactly like ordinary punctuation and are not. GPT models write the non-breaking hyphen in compound words, and it breaks search and find-and-replace like any other hidden character.
Replaced with a normal hyphenU+2011Non-breaking hyphenFormatting controls that set text direction. Harmless to read, obvious to anything inspecting the underlying characters, and never intentional in ordinary prose.
Removed completelyU+200ELeft-to-right markU+200FRight-to-left markU+061CArabic letter markU+2066Left-to-right isolateU+2067Right-to-left isolateU+2068First strong isolateU+2069Pop directional isolateU+202ALeft-to-right embeddingU+202BRight-to-left embeddingU+202CPop directional formattingU+202DLeft-to-right overrideU+202ERight-to-left overrideFamilies used to smuggle invisible payloads into text. A tag-character sequence can encode an entire hidden message that renders as nothing, so these are removed wholesale rather than listed one by one.
Removed completelyU+FE00 to U+FE0FVariation selectors 16U+E0100 to U+E01EFVariation selectors supplement 240U+E0000 to U+E007FTag characters 128This is where a Claude watermark remover that only strips characters leaves you exactly where you started.
It’s worth notingU+200B that while there are several factors to consider here, the approach that tends to work bestU+00A0 is one that balances both competing priorities.
It’s worth noting that while there are several factors to consider here, the approach that tends to work best is one that balances both competing priorities.
Every invisible character is gone, and the four hedges that read as Claude to a person, and as predictable to a detector, are flagged for the rewrite.Claude hedges more than any other major model. It qualifies before it commits, it stacks clauses inside a single sentence, and it reaches for paired em dashes to hold them together. None of that is hidden, which is precisely why it survives every character cleaner ever written.
The detector reads the writing as well as the bytes, quoting each flagged phrase back with the pattern it matched so you can see what is giving the draft away. Detection costs nothing. Full rewriting is the paid part.
Copy straight out of Claude and paste it in. Do not reformat first, because that can strip the very characters the Claude watermark detector is looking for.
Two passes run at once. One reports every hidden character by Unicode code point and count. The other reads the writing and flags Claude's hedging, qualifying clauses, and stock transitions.
Each finding shows what it is and why it was flagged. Nothing is rewritten behind your back.
The Claude watermark remover applies only what you accepted. Your meaning, names, and structure stay as they were.
Three different things now share the name. First, Anthropic's real statistical watermark, live since 14 August 2026, which lives in word choices. Second, the invisible Unicode characters that travel with copied Claude text: zero-width spaces, non-breaking spaces, word joiners, formatting attributes. Third, the phrasing fingerprint, which for Claude means heavy hedging and long qualifying clauses. GPTCleanup removes the hidden characters and finds the AI-sounding phrasing, then rewrites it so the draft reads naturally.
Yes, since 14 August 2026. Anthropic ships a statistical watermark in new Claude models, applied at the model level and switched on globally rather than only in the EU. It works by steering low-stakes word choices with a key, leaving a pattern across the passage that a reader cannot see. It is the same family of approach as Google DeepMind's SynthID-Text. Older Claude models are being covered over the following months.
It is the reason it happened. The EU AI Act's transparency code took effect on 2 August 2026 and requires AI-generated content to be machine-identifiable. Anthropic responded with text watermarking and chose to apply it worldwide rather than geofence it.
It scans your draft twice. The character pass compares every character against 421 supported characters and reports each match by Unicode code point with a count. The phrasing pass reads the writing itself and flags the hedges, qualifying clauses, filler transitions, and punctuation habits Claude falls back on, quoting each one with the pattern it matched.
Stripping invisible characters from your own text is ordinary formatting and it is legal. Anthropic's statistical watermark is a different matter: it exists to satisfy EU AI Act transparency rules, so deliberately defeating it runs against a disclosure regime rather than a formatting preference. Whatever you do with the text, platform terms and any academic or workplace policy on AI assistance still apply.
It is the most recognisable thing about Claude's prose and none of it is hidden. Long qualifying clauses, paired em dashes, and hedges like “it is worth noting” appear far more often than in human writing. Because they are real punctuation and real words, no character cleaner touches them. They are flagged by the phrasing pass instead, and you decide what to change.
Removing hidden characters alone does not change how a detector reads a draft: ordinary AI detectors score phrasing and sentence rhythm, not invisible characters. That is why GPTCleanup also scans for the AI-sounding patterns and giveaways and rewrites them so your writing reads naturally. No tool can promise a result against any specific detector.
No. Word processors, CMS editors, PDF copy-paste, and translation tools all leave the same characters behind. Their presence tells you the text passed through software, not which software and not who wrote it.
No. Zero-width and directional characters render as nothing, so removing them cannot alter a word. Look-alike spaces become ordinary spaces, which keeps words apart rather than joining them. Phrase rewrites are shown in context and applied only where you accept them.
Yes. Detection is free and needs nothing but your text. Full phrase-level rewriting is the paid part of GPTCleanup, and it is a separate job from stripping characters.
They are not all in the same position, which is why each has its own page. Google is the outlier: SynthID genuinely watermarks Gemini text by adjusting word probabilities during generation, so no character cleaner can reach it. OpenAI and xAI have shipped nothing comparable for text, so what turns up in ChatGPT and Grok output is interface residue like Claude’s.
The same hidden-character set applies across all of them. What differs is the phrasing, and that is what each page covers in that model’s own voice: ChatGPT watermark remover, Gemini watermark remover, and Grok watermark remover.