Claude watermark cleaner
What the Claude watermark cleaner does
Paste text, press Scan text to see what is in it, press Clean it to take it out. The cleaner finds 14 specific invisible Unicode characters, deletes eleven of them and swaps the other three for an ordinary space. It also squeezes runs of two or more spaces or tabs inside a line down to one and trims the ends of lines.
It runs in your browser. Nothing is uploaded, there is no account, no queue and no cap on how long your text can be or how many times you can run it. Once cleaned, a badge tells you how many characters came out and a copy button puts the result on your clipboard.
Read this first
Claude's text watermark is not a character
Anthropic's own explanation of how Claude's text watermark works says it plainly: nothing is added to the text and there are no hidden characters. It is a pattern in Claude's word choices, a version of the SynthID-Text approach published by Google DeepMind, and it is read with a key rather than found by searching.
So this tool does not remove it. No character-stripping tool does, because there is no character there to strip. If you find a page claiming it removes Anthropic's watermark by scanning for invisible characters, it is describing a different thing entirely. Anthropic's own note says light editing probably will not remove the watermark and a complete rewrite where every word is replaced will.
What this tool removes is the other thing: the real invisible characters that end up in text you copied out of a chat window. Those are separate from the watermark, they cause real problems, and they are what a cleaner can actually fix.
How to
How to use the Claude watermark cleaner in three steps
Scan before you clean. The report tells you what is actually in the text, which is often nothing, and that is worth knowing before you change anything.
- Step 1
- Paste your text into the box. There is no length limit. If you just want to see what a dirty paste looks like, press Try a sample.
- Step 2
- Press Scan text. Each finding is marked in the report with the reason on hover, and the two counters show hidden Unicode characters and flagged phrases.
- Step 3
- Press Clean it. The badge says how many hidden characters were removed, or that none were found. Then press Copy.
The list
The 14 characters it handles
Eleven are deleted outright. Three look like spaces, so they are replaced with a normal space instead of removed, because deleting them would join two words together.
- Zero-width family
- Zero-width space, zero-width non-joiner, zero-width joiner and word joiner. Invisible, they take up a character slot, and they survive copy and paste.
- Byte-order mark
- A byte-order mark stranded in the middle of a document confuses parsers, feeds and validators, and shows as a stray glyph in some editors.
- Soft hyphen
- Invisible until a line wraps, at which point a hyphen appears in a word that never had one.
- Space look-alikes
- No-break space, narrow no-break space and hair space. These are the three swapped for a normal space rather than deleted.
- Directional controls
- Five bidirectional formatting characters: two embeddings, a pop, and the left-to-right and right-to-left overrides. They can reorder how text is displayed.
- Emoji are safe
- A zero-width joiner sitting between two emoji is what builds a single composed glyph. The cleaner leaves those alone and only removes stray ones.
The cause
Where the invisible characters come from
If the watermark adds none, why is your paste full of them? Because they come from the plumbing, not the model. Web chat interfaces render text with formatting characters in it. Copying from a rendered page rather than a plain text field brings those along. Word processors and note apps insert no-break spaces around numbers and dashes.
That is why a scan often finds nothing in one paste and several characters in another from the same conversation. It depends on where you copied from and what the intermediate app did, not on which assistant produced the words.
Consequences
What hidden characters actually break
They break exact matching, which is most of what software does. Find and replace misses a word that has a zero-width space in it. A URL slug with one in it resolves to nothing. A spreadsheet lookup fails against a value that looks identical on screen.
Developers hit it hardest. A git diff shows a line as changed when nothing visible changed. A linter reports an error at a column with nothing in it. An SQL insert stores a value that never matches on the way back out. None of these look like a character problem until you go looking for one.
The blunt question
Does the Claude watermark cleaner make text undetectable?
No. Cleaning invisible characters changes nothing about how the words are chosen, and word choice is where Claude's watermark lives. It also does nothing to the patterns a detector reads. We make no promise about any detector's result and do not test against one.
There is a separate point worth making. Removing a machine-readable mark does not remove a disclosure duty. If your school, your client or a platform requires you to say that a text was drafted with an assistant, that requirement stands whatever a cleaner did to the characters.
In practice
Who cleans text, and when
Developers pasting Claude output into code, config or a commit message. Writers moving drafts between a chat window, a document and a publishing system. Anyone whose search or find and replace has started failing on text that looks perfectly normal.
The habit that helps is scanning before pasting into anything that does exact matching: a repository, a database, a CMS, a spreadsheet. It costs a second, it is free, and the text never leaves your browser.