The output
What the Llama watermark cleaner gives you
Two things, in order. A scan that counts the invisible Unicode characters in your text and the AI-sounding phrases in it, with a reason attached to every item. Then, if you want it, a cleaned version with those characters removed and a badge stating exactly how many came out.
The two halves are deliberately separate. The scan is information and changes nothing. The clean is an edit and changes only what it says it changes: invisible characters and repeated spacing. No word is rewritten, which means you can run this over finished copy without reviewing it again afterwards.
Provenance
Where invisible characters get into a document
Rarely from one place. By the time text reaches a CMS it has usually been through a chat window, a document editor and a clipboard, and each of those can leave something behind. Treating this as a model problem is a mistake that sends people looking in the wrong place.
The practical consequence is that a scan is worth running on any text with a copy-paste history, not only on model output.
- Chat interfaces
- Zero-width spaces and hair spaces turn up in copied replies, sometimes as formatting artefacts rather than anything deliberate.
- Word processors
- No-break spaces around numbers and units are inserted automatically and look exactly like ordinary spaces.
- PDFs and slide decks
- Copying out of a rendered layout collects soft hyphens and directional controls that were doing layout work.
- Anything multilingual
- Directional marks are added around mixed left-to-right and right-to-left runs and get stranded when the text is edited.
Workflow
How to clean a Llama watermark in three steps
All three steps run in the page. Nothing is uploaded, so none of this depends on a connection once the page has loaded.
- Step 1
- Paste the text. There is no character limit here because there is no server on the other end of it.
- Step 2
- Scan, and read the report before you act. Each finding carries the reason that character matters.
- Step 3
- Clean, check the badge for the count removed, and copy the result back into your document.
Sequencing
Scan before you clean, every time
It is tempting to skip straight to cleaning, and mostly it would be fine. The reason not to is that the scan tells you where a problem came from, and that changes what you do next. Fourteen zero-width spaces evenly distributed through a paragraph is a different story from one no-break space in a date.
The scan is also the only record you get. Once the text is cleaned, the evidence is gone by design, so if you need to tell a colleague why their export was breaking, read the report first.
Careful cases
Cleaning text bound for code, CSV or a CMS
Prose is the easy case. Structured text needs one look before you paste the result back, because the cleaner normalises spacing as well as removing characters.
None of these are reasons to avoid the tool. They are reasons to look at the output for ten seconds rather than pasting it straight into production.
- Code
- Runs of two or more spaces collapse to one, which can disturb alignment you meant to keep. Indentation at the start of a line is not touched.
- CSV and JSON
- A stray byte-order mark is exactly the kind of thing that breaks a parser, so this is the highest value case for a scan.
- Persian and Indic scripts
- The zero-width non-joiner is meaningful there. Review such text before cleaning rather than after.
- Emoji
- Safe. Joiners inside emoji sequences are left in place, so multi-part emoji do not fall apart.
Scope
What the Llama watermark cleaner will not do
It will not rewrite anything. The AI-sounding phrases the scan counts are reported for your attention and left exactly as they are, so this is not a humanizer and does not pretend to be one.
It will not remove a statistical watermark either. A signal embedded in which words a model chose sits across the document as a whole, which is why paraphrasing a sentence or two barely moves it, and why a tool that changes no words cannot affect it at all. And it makes no promise about any detector, because detectors read the writing rather than the characters.
Privacy
Local processing, and what that means in practice
The scan and the clean both run in your browser. No text is transmitted, no account exists and no run counter applies, which is why this page carries no allowance while the analysis tools on this site do.
For a workflow the useful consequence is that there is no sensitivity threshold to think about. Client copy under NDA, unpublished material, internal documents: none of it leaves the machine, so the decision about whether it is safe to paste does not arise.
In practice
Who runs the Llama watermark cleaner, and when
Publishing teams who found a slug with a character in it that nobody typed. Developers debugging a string comparison that fails on two lines that look identical. Editors taking delivery of copy that has been through three tools and a shared document.
Most people arrive after the symptom, not before it. The scan is quick enough that it is worth making it the first check rather than the last, because ruling out invisible characters in one second saves an hour of looking at the wrong thing.