Does the ChatGPT watermark survive copy-paste, Word and Docs?

Pasting a ChatGPT answer into Word or Google Docs should not take the watermark out, when the answer carries one, by our reading of how it works. Neither program documents a way to show it either. What they can show is a different thing, formatting marks, and the two get confused because both are invisible. This article separates them and gives the steps for each program, read from Microsoft’s and Google’s own help.

Does copying ChatGPT text keep the watermark?

The ChatGPT watermark is carried by the words, so copying the words copies the mark. OpenAI announced it on 5 October 2026 for eligible ChatGPT and Codex text in the European Union, rolling out over the coming weeks, and as an opt-in for select models in the API. Its technical report describes a secret key that steers which next word the model picks, and a detector with the same key tests the finished text for that lean. A clipboard moves letters and spaces, and the signal consists of which words were chosen.

The copy-and-paste claim comes from the press and not from OpenAI’s post, which does not discuss copying. TechCrunch wrote that because the mark lives in the words themselves, “it travels with the text when it’s copied and pasted.” That is a reasoned statement about the design, and OpenAI has published no test of word processors. Scope matters more than the clipboard: text outside the announced scope, from before the rollout reached an account, or from an API call without the option switched on is not described as watermarked by OpenAI.

What can Word and Google Docs show, and what can they not?

Neither Word nor Google Docs can display the ChatGPT watermark, and no help page from Microsoft or Google describes a feature that could. That is a deduction from three facts. The mark is a statistical pattern across many word choices, so no single character carries it. The detector needs OpenAI’s secret key. And OpenAI is giving detector access only to approved researchers and expert organisations at first, citing the risk of missed watermarks and false positives.

A word processor can do two nearby things. It can show formatting marks, and that is usually what people find when they search “how to see ChatGPT watermark in Word”. And it can hold invisible characters pasted in from elsewhere. Neither is textGrain. OpenAI’s post and report describe the signal as word choice and do not mention inserting a character.

How do you show formatting marks in Word and Google Docs?

Each program has one switch for its formatting marks, and each vendor documents it. The steps below follow the help pages as read on 12 October 2026.

• Word on Windows: press the Show/Hide button, which Microsoft says turns formatting marks such as spaces, paragraph markers and tab marks on and off. The shortcut is Ctrl+Shift+8 on the main keyboard, and Microsoft’s shortcut list says not to use the numeric keypad.
• Word, always-on marks: go to File, Options, Display and tick the marks you want under Always show these formatting marks on the screen. Microsoft says they stay visible whether or not Show/Hide is on.
• Word on Mac: go to Word, Preferences, View and tick the marks under Show Non-Printing Characters. Microsoft’s shortcut page has no Mac row for this command.
• Google Docs: open View, then choose Show non-printing characters. Google says the marks appear in blue, and that choosing the option again hides them.

Neither list names a zero-width character, and neither page says whether such a character would appear. To check for one, use our invisible character detector. It reports exactly eight characters, among them the zero-width space, the byte-order mark and the soft hyphen, so a clean result means only that. Our zero-width guide gives the Word search code. The narrow no-break space (U+202F) is not among the eight. Rumi reported it in some 2025 ChatGPT answers, where it looks like a regular space in ordinary word processors, quoted OpenAI calling it “a quirk of large-scale reinforcement learning”, not a watermark, and later wrote that it no longer saw them. To find it, search for its code in your editor.

Do fonts, formatting or a PDF export change the signal?

Fonts, styles and file formats should leave the watermark untouched, and this is our inference and not a published result. OpenAI places the signal in which words the model chose. A font changes how a word looks, a style changes its weight or size, and an export to PDF or another format stores the same words. OpenAI has not tested these cases, and a conversion tool that rewrites or translates as it goes is a different matter, because it changes the words.

What weakens the signal, and why is that not a method?

Changing the words weakens the signal. In OpenAI’s evaluation on English answers to ELI5 questions, in 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. Our article on what textGrain is covers the detector’s other limits.

Those figures describe the limits of a detector, and they do not amount to a procedure. A passage edited that far is a passage whose wording someone has changed, and a lower detection rate is not proof of anything. OpenAI puts it directly: “The absence of a detected watermark does not prove human authorship.” Our rewriter replaces wording and shows each change. We make no claim about what a keyed detector would find, and no public tool can test that.

Who can check a text for the watermark today?

Only OpenAI’s detector can, and at launch it is open to approved researchers and expert organisations. No word processor, browser extension or character scanner reads the mark, ours included.