ChatGPT text now carries a watermark
On October 5, 2026, OpenAI announced textGrain, an invisible watermark for text written by ChatGPT and Codex. It is rolling out to every ChatGPT plan in the European Union over the following weeks, and developers anywhere can switch it on for their own apps through the API. The move answers the EU AI Act rule that AI-generated content must be machine-identifiable.
textGrain is not a hidden character you can find and delete. It is a statistical pattern in word choice: whenever several words would fit a sentence equally well, a secret key nudges the model toward one of them. A reader sees ordinary text. A detector holding the key sees that too many of those choices line up.
Because the pattern lives in the words themselves, it survives copy and paste, a trip through Word or Google Docs, and any change of font or formatting. Google does something similar with SynthID for Gemini text, so the same applies to drafts from other assistants.
What removes it and what does not
OpenAI published its own numbers on how the signal holds up. On a clean, unedited answer of about 200 tokens the detector finds the watermark roughly 80% of the time, rising to about 95% at 400 tokens. Swap 10% of the words for synonyms and detection drops from around 92% to 66%. Swap 25% and it falls to about 17%. Translation and heavier editing weaken it further.
- Does not remove it: pasting into Notepad, clearing formatting, changing fonts, or running a character cleaner. None of these touch the words.
- Barely dents it: fixing a few words by hand or swapping the odd synonym. Most of the original choices are still there.
- Removes it: rewriting every sentence so the wording, order and rhythm are new while the meaning stays the same.
The Humanizer does the last one. Every sentence is re-drafted from scratch, so far more than a quarter of the words change. On three ChatGPT drafts we ran through Enhanced mode, 65% to 87% of all three-word sequences were new in the result, while names, numbers and the order of the argument stayed put.
How to remove the watermark from ChatGPT text
- Paste your ChatGPT, Codex or Gemini text into the editor above. Free runs take up to 200 words.
- Choose the purpose: Essay, Email, Blog/Article or General.
- Keep the mode on Enhanced and click Humanize.
- Read the result and its AI detector scores, then copy it.
For long documents, paid plans rewrite up to 3,000 words in one run. See pricing.
Hidden characters are a separate thing
Before textGrain, the "ChatGPT watermark" people talked about was a different story. In April 2025, users found narrow no-break spaces (Unicode U+202F) and other invisible characters in answers from some OpenAI models. OpenAI said these were a quirk of training, not a watermark, and they later stopped appearing. Zero-width spaces and non-breaking spaces still turn up in text copied out of chat windows.
Those characters are easy to strip and they matter for practical reasons: they break word counts, search and replace, and form fields. The free AI text cleaner removes them, along with smart quotes, long dashes and leftover markdown, in one click. The humanizer output is plain text already, so if you rewrite, you do not need to clean afterwards.
Watermarks and AI detectors are different checks
A watermark check needs OpenAI's secret key, and for now only approved researchers and expert organizations can run it. The checkers teachers, editors and clients actually use, such as Turnitin, GPTZero and ZeroGPT, never look for a watermark. They score writing style: how predictable each word is and how evenly the sentences run.
That is why the rewrite does both jobs at once. The new wording breaks the watermark pattern, and the uneven, natural rhythm is what style-based detectors read as human. You see the detector scores for the result right under it, so you know where the text stands before you send it. More on each detector: Turnitin, GPTZero, ZeroGPT.