AI-written text often comes with characters nobody typed, and this guide from AmICited (FlowHunt's free AI brand-visibility tracker) shows how to clean them out with the free LLM Watermark Remover before the text goes live. We explain what the remover takes out and what it leaves, the two steps it runs (a deterministic cleanup, then a model rewrite), how to check the result, when a cleanup is the right call, why text watermarks come off so easily, and why no tool can honestly promise that every trace of AI is gone.
The short version: paste the text, get back a cleaned and rewritten draft, read it, and check it with the LLM Watermark Detector if it matters. The result is a draft for you to approve, not a finished page.
What a ChatGPT watermark remover actually removes
AI text can be marked in two ways. Character watermarks swap ordinary characters for lookalikes: a normal space for a three-per-em space (U+2004) or an ideographic space (U+3000), a Latin letter for its Cyrillic twin, or invisible characters such as zero-width spaces slipped between words. Hidden characters also arrive for plain reasons, from the chat interface or the editor the text passed through. Either way they are harmless to a reader and a nuisance to search engines, spell checkers, content systems and anything that compares text exactly.
Statistical watermarks, such as Google's SynthID Text for Gemini, work differently. The model picks, among the few most likely next words, the one that scores highest under a secret scheme, so the mark lives in the word choices rather than in extra characters. There is nothing to strip, and only the provider's own detector can say whether the pattern is still there.
Why text watermarks come off so easily
The engineer Sean Goedecke makes the case bluntly in his write-up on the subject: text watermarks will always be trivial to remove. A character watermark comes off by replacing each lookalike with the ordinary character it imitates, which is a rule a program can apply in milliseconds. A statistical watermark comes off by rewording, because once the words change, the pattern in the word choices changes with them. Paraphrasing with a different model, or editing by hand, is enough.
Signed metadata does not close the gap either. C2PA Content Credentials can prove where a file came from, but they attach to files. Text copied out of a chat window is plain characters with nowhere to carry a signature.
That is why this tool runs two steps: a rule-based cleanup for the characters, and a rewrite for the wording.
Step one: the deterministic cleanup
The first pass edits characters by rule. The same input always gives the same output, and every change is recorded, so you can see what was removed and where.
- Hidden characters are removed: zero-width spaces (U+200B), zero-width non-joiners and joiners (U+200C, U+200D), word joiners (U+2060) and stray byte order marks (U+FEFF).
- Bidirectional controls that can reorder how a line displays, such as U+200E, U+200F and U+202A to U+202E, are removed.
- Invisible Unicode tag characters (U+E0000 to U+E007F) are removed.
- Unusual spaces, such as the narrow no-break space (U+202F), are turned into ordinary spaces.
- Lookalike letters, such as a Cyrillic a standing in for a Latin one, are not swapped by this pass. They are reported, and the rewrite in step two usually replaces them.
Step two: the rewrite
After the cleanup, the text is sent to a language model to be rewritten: the same meaning, reworded. This is what catches what a rule cannot, such as lookalike letters inside words, and it also varies the phrasing. It is a best effort, not a guarantee, and because it rewrites, it can shift a nuance, a number or a claim.
Your text goes to the configured model provider for this step. Do not paste anything you would not send to an AI service, such as personal data, credentials or confidential drafts.
A worked example
Here is a sentence with two invisible characters, what the cleanup leaves, and a rewrite of the kind the second step returns.
Before: The results⟦U+200B⟧speak for⟦U+202F⟧themselves. Cleanup: The results speak for themselves. Rewrite: The numbers make the case on their own.
When to clean text, and when not to
Cleaning is the right call when invisible characters would cause real harm: copy headed for a web page, a product feed, a CMS import or code, where one hidden character can break search, matching or formatting. It is also a sensible last step for any AI-assisted draft you have edited, checked and now stand behind.
It is the wrong call when the goal is to pass off AI text as something it is not. The EU AI Act places the marking duty on AI providers, but it also expects anyone who publishes AI-generated text to inform the public on matters of public interest to say so, unless a person has reviewed it and taken editorial responsibility. Schools, publishers and employers set their own rules too. Clean text for quality, and be open about how it was written where it matters.
Check the result, then get found
Always read the rewrite before publishing it, and run the LLM Watermark Detector on the output if hidden characters are a concern. What matters after that is whether AI assistants recommend what you published. AmICited asks ChatGPT, Perplexity and Google AI about your brand on a schedule and tracks how the answers change, so you can see whether clean, well-structured content turns into mentions.
To see where you stand today, run the free AI Brand Visibility Report on your brand.
Frequently asked questions
Why are text watermarks so easy to remove?
Because text has almost nowhere to hide a mark. A character watermark is undone by swapping each lookalike back to the ordinary character, and a statistical watermark, which lives in word choice, fades as soon as the text is reworded. Signed metadata such as C2PA does not help, because text copied from a chat window has no file to carry it.
Can I remove the ChatGPT watermark from text?
You can remove the hidden and lookalike characters that usually get called a ChatGPT watermark: the cleanup strips invisible characters and turns unusual spaces into ordinary ones, and the rewrite rewords the text. Rewording is also what weakens a statistical watermark, but only the provider's own detector can confirm one is gone, so this tool does not claim to.
Does the remover change the meaning of my text?
The cleanup step does not: it only removes or normalizes invisible characters. The rewrite step rewords the text and aims to keep its meaning, but a rewrite can shift a nuance, a figure or a claim. Read the result before you use it.
Is removal guaranteed?
No. The cleanup removes the characters it recognizes, and the rewrite catches much of the rest, such as lookalike letters. It is a best effort. If it matters, run the LLM Watermark Detector on the output to confirm what is left.
Where does my text go?
The cleanup runs by rule. For the rewrite, your text is sent to the configured language model provider. Do not paste personal data, credentials or anything confidential.
Will this make AI text undetectable?
That is not what it is for, and it cannot promise it. AI detectors used for authorship look at statistical patterns in the writing, not only at hidden characters. Use the remover to clean text for publishing, and follow the AI-use rules of your school, publisher or employer.
What is the difference between the detector and the remover?
The LLM Watermark Detector only reads text and reports the hidden characters it finds, with their positions. The LLM Watermark Remover changes the text: it strips those characters and returns a rewritten draft. Run the detector first to see what is there, and again after the remover to check the result.
Next, see what AI says about your brand
AmICited asks ChatGPT, Perplexity and Google AI about your brand on a schedule and tracks how the answers change.