How to Remove a ChatGPT Watermark from Text

ChatGPT does not officially watermark text. No public OpenAI documentation describes a deliberate text watermark, and nobody outside the company can confirm one either way. What ChatGPT output does contain, reliably, is characters that do not belong in ordinary typing.

The narrow no-break space is the usual example. It looks exactly like a space, sits at codepoint U+202F instead of the ordinary U+0020, and survives copy and paste into almost anything. Zero width characters behave the same way and take up no room at all. Whether anybody meant these as a fingerprint is beside the point. They work like one, because they are trivially greppable and no human types them by hand.

You are not removing a watermark that OpenAI admits to. You are removing characters. Characters can be counted, which is what makes removing them a solved problem.

The hidden characters fall into eight families, 104 types in total, of which 28 render as literally nothing on screen. Most guides name one or two and stop. The families matter more than the individual codepoints, because each family needs different handling.

The split in that last column is the important part. Deleting an invisible character costs you nothing. Rewriting a visible one changes your text, so it stays your decision rather than the tool's. One family in that table earns a section of its own, because tag characters do something none of the others can.

Tag characters matter more than zero width spaces because they carry a readable message. A zero width space is one invisible bit. The tag block, U+E0000 to U+E007F, maps one to one onto printable ASCII, so a run of tag characters is a whole sentence hiding inside text that looks completely normal.

Subtract 0xE0000 from a tag character and you get an ASCII code. That is the entire trick. U+E0041 is the letter A. String enough of them together and you can hide an instruction, a name, or a tracking string in a paragraph nobody can see it in.

This is why the scanner decodes rather than silently deleting. We checked the three cleaners ranking for this query in August 2026. All three strip the characters and report a count. One can show you where they sit. None reads them back. Being told your text was carrying the words "ignore all previous instructions" is a different piece of information, and it is the one you want before pasting that text into another AI system.

Checking for them is the same operation as removing them, which is worth knowing before you decide you only want to look.

Checking whether text has a ChatGPT watermark means counting characters that should not be there. Paste it in and read the scan panel without pressing Clean. You get each codepoint, how many times it appears, and a preview that marks every one in position. Nothing is modified until you ask.

Two manual checks work if you would rather not trust a tool. In Microsoft Word, turn on formatting marks and unusual spaces show up as odd symbols rather than gaps. Any code editor does the same with "render whitespace" switched on. Both miss tag characters and lookalike letters entirely. That gap is the reason the scanner exists.

Word is also where most people hit this problem, so it is worth covering on its own.

Removing ChatGPT watermarks in Word or Google Docs works the same way as anywhere else, with one extra trap: pasting back in can reintroduce formatting you just stripped. Clean the text first, then paste it as plain text.

• Word: copy your text out, clean it, then paste back with Paste Special and choose Unformatted Text. Pressing Ctrl and Shift and V does the same thing in most versions.
• Google Docs: use Edit, then Paste without formatting. The keyboard route is Ctrl or Cmd, Shift and V.
• Either: turn on formatting marks first and you can see the problem before and after, which is a faster check than trusting the paste worked.

The popular Notepad trick does less than people think. Notepad strips formatting, not Unicode. Zero width spaces and tag characters are perfectly valid text and they come through untouched. Plain text is not the same as clean text, and that gap is where most of what a remover cannot do begins.

Does ChatGPT put a watermark in text?

OpenAI has never confirmed a deliberate text watermark. ChatGPT output does routinely contain non-standard characters such as the narrow no-break space, which behave like a fingerprint whether or not anyone intended them to.

Is removing a watermark from AI text legal?

Stripping invisible characters from your own text is not illegal. It can still breach an academic integrity policy or a platform term of service, and that is a separate question from legality. Check the rules that apply to you.

Will removing hidden characters stop an AI detector flagging my text?

No. Detectors judge word choice and sentence structure, not invisible characters. Cleaning the characters fixes formatting and stops trivial fingerprinting. It does not change how the writing reads.

Does pasting into Notepad remove a ChatGPT watermark?

Partly. Notepad strips formatting and some control characters, but it preserves most invisible Unicode, including zero width spaces and tag characters, because they are valid text. Plain text is not the same as clean text.

Do other AI models leave the same characters?

The character families overlap but the mix differs by model. Claude output tends to carry a different profile, which is why it has a separate tool.