In the spring of 2025, people examining output from OpenAI's newer models noticed something odd: the text was sprinkled with characters that looked like spaces but weren't. The most common was U+202F, the narrow no-break space. The discovery set off a wave of "ChatGPT is watermarking your essays" posts, and an ecosystem of "watermark remover" web tools appeared almost overnight.
The launch is set for 9:41 AM on Friday.
the same line, revealedThe launch is setU+202Ffor 9:41NBSPAM on Friday.
So: does ChatGPT watermark its text? The honest answer has three parts.
The characters are real
This part is not in dispute. Model output has repeatedly been observed carrying invisible or near-invisible Unicode: narrow no-break spaces where a keyboard space belongs, non-breaking spaces, occasionally other format characters. Paste that text into a code editor with invisibles rendering (or any of these five methods) and you can see them yourself. They survive copy-paste, they survive most "paste as plain text" commands, and they end up in documents, CMSs and codebases.
"Watermark" is probably the wrong word
A deliberate text watermark, in the research sense, does not need funny characters at all. Statistical watermarking (the approach in Google's SynthID Text, and in OpenAI's own published research) biases which ordinary words the model picks, embedding a signature in word choice itself that a detector can later test for. Nothing about it is visible OR invisible at the character level.
Character-level oddities, by contrast, look like what forensics folk concluded they were: training artifacts. The models learned from professionally typeset text - academic publishing, multilingual typography - where narrow no-break spaces are correct punctuation (French puts one before every colon and question mark). The model reproduces the typography of its best sources. OpenAI said as much when the story broke, and later model updates reduced the behavior.
The distinction matters less than it sounds like it should. Whether the characters are a fingerprint or an accident, the practical effect on you is identical: your paste contains characters you did not choose, that most people cannot see, that mark the text as machine-touched, and that break search, validation and code in the usual ways.
What this means if you use AI tools
If you paste AI-assisted text into anything that matters, assume the invisible layer exists:
- Typography tells travel too. Beyond invisibles, model output leans on curly quotes, real ellipsis characters and the infamous em dash - visible tells that read as "pasted from a bot" to a growing number of people.
- The fix is not a website. Pasting your text into a random "watermark remover" page means handing your text to a stranger to solve a privacy problem. Read the room.
- Clean at the clipboard. A local, automatic cleaner catches every copy without a manual step and without uploading anything. That is precisely what CopyClean is for: it strips the invisible junk and normalizes the AI typography on-device, the instant you copy.
One thing we will not tell you: that removing these characters makes AI text "undetectable". It does not, and we wrote a whole post about why. Clean text is about quality, privacy and not shipping invisible bugs - not about fooling anyone.