How to remove AI watermarks from your text
"Remove the AI watermark" means two different jobs depending on what actually put it there. One is a two-second fix. The other cannot be done by any tool, including this one, and the rest of this guide explains why before it wastes your time.
To remove a watermark from AI-generated text: paste it into a checker that exposes hidden Unicode characters, review what it finds, then copy the cleaned version. That deals with zero-width spaces, tag characters and the other invisible debris chat interfaces leave behind. It does nothing to a statistical watermark such as Google's SynthID or Anthropic's confirmed mark for newer Claude models, because that kind of mark lives in which words the model chose, not in any character you can select and delete.
More: what was found, x-ray view and options
X-ray view
Your text with every hidden character exposed as a labelled chip. Hover a chip for its Unicode name.
What was found
| Code point | Character | Count | Action |
|---|
Step 1: find out which kind of watermark you're dealing with
Before cleaning anything, check the vendor. As of this guide's review date, two providers have a confirmed, deployed text watermark: Claude, for models released after 2 August 2026, and Gemini, via SynthID. ChatGPT, Copilot, Grok, DeepSeek, Perplexity, Mistral and Llama do not have one deployed yet, though several have committed to adding one under the EU's transparency rules.
This matters because the two watermark types have opposite properties, and confusing them is where most bad advice on this topic comes from:
| Hidden characters | Statistical watermark | |
|---|---|---|
| What carries it | Extra invisible characters in the text | Which words the model picked |
| Can you see it | Yes, in any hex dump or x-ray view | No, not even in principle |
| Removable by cleaning | Yes, completely | No, there is nothing to delete |
| Deployed by major providers | No provider uses this as a watermark | Yes, Google and Anthropic so far |
Full mechanism, if you want it: statistical versus character watermarks.
Step 2: clean the part that's actually there
Whatever the vendor, text copied out of a chat interface usually carries invisible characters that have nothing to do with watermarking, non-breaking spaces holding units to numbers, zero-width spaces at line-break points, soft hyphens from justified text. Paste your text into the checker above and it will:
- Show every hidden character in an x-ray view, labelled by what it is and how risky it is.
- Tally them by code point, so you know exactly what you're dealing with before changing anything.
- Produce a cleaned copy in the same pass, ready to paste elsewhere.
It runs entirely in your browser. Nothing you paste is uploaded, which matters if the text is a client email or a snippet from a private repository.
Step 3: know what didn't just get fixed
If the text came from Claude or Gemini, the cleaned copy you just produced still carries the statistical mark, unchanged. Nothing in step 2 touched it, because nothing could. The signal is the pattern of word choices across the passage, and every word is still exactly the word it was.
The only thing that disturbs a statistical watermark is changing enough of the actual words, which means rewriting or substantially paraphrasing the passage, not cleaning it. That's a different operation with a different set of trade-offs (meaning drift, tone changes, the risk of introducing errors), and it's outside what a character cleaner does or should attempt. This site does not offer it.
Be skeptical of anyone who claims otherwise. A tool that says it can strip Google's or Anthropic's watermark in your browser is either doing something else, usually feeding your text through a paraphraser and calling it "watermark removal," or it doesn't work. Neither statistical watermark has a public detection or removal method, by design.
Common situations
"My text won't paste into this form correctly"
Almost always a hidden-character problem, not a watermark one. A no-break space or zero-width character is breaking validation. Clean it and the form should accept it.
"A CSV import keeps failing on one row"
Same cause. A stray invisible character turned a numeric field into text, or shifted a delimiter. The hidden characters guide covers the usual sources.
"I need to paste AI output into source code"
Clean it first. Bidirectional control characters can make code display in an order different from how it executes. See invisible characters in code for what that looks like.
"I want to know if this text was written by AI"
Cleaning won't tell you that, and style-based "AI detectors" are unreliable enough to produce false positives on ordinary human writing. That's a different problem than this guide solves.
Frequently asked questions
Can I remove an AI watermark from text for free?
The hidden-character kind, yes, completely and for free, with no sign-up. The statistical kind that Claude and Gemini now use cannot be removed by any tool, free or paid, because there's no character to delete.
Does removing hidden characters change my wording?
No. Every character removed is invisible to begin with, and unusual spaces are replaced with ordinary ones rather than deleted, so words never run together. The visible text and word order are untouched.
How do I remove an AI watermark from code?
The same way as text: paste it into the checker. It preserves indentation, tabs and line endings exactly, and flags bidirectional control characters in red since those are the ones that can make code misleading rather than merely messy.
Will this work on ChatGPT, Claude, Gemini or Copilot output?
The cleaning step works identically on text from any of them, hidden characters aren't vendor-specific. Whether there's also a statistical mark to worry about depends on the vendor; the per-vendor pages linked in step 1 above cover the current status of each.
Why do some invisible characters stay after cleaning?
Because they're not debris. A zero-width joiner holding a family emoji together, a zero-width non-joiner required in Persian or Indic scripts, a blank braille cell, those are content, not noise, and the cleaner checks context before deciding.
Related reading
- The AI watermark remover and checker - paste text or code and see this in action.
- Can an AI watermark be removed? - the short honest answer: what comes out completely, what stays, and why.
- Statistical versus character watermarks - the full mechanism behind why one survives cleaning and the other never existed as a watermark.
- What the EU AI Act requires - the regulation driving vendors toward watermarking in the first place.
- Hidden characters in text - where the debris actually comes from.