OpenAI now embeds an invisible watermark, called textGrain, in the text ChatGPT and Codex generate. It's applied automatically across every plan in the EU. OpenAI's API customers anywhere in the world can enable it in their output. The ChatGPT Watermark Remover strips it out at the sentence level, not with a synonym swap, so your words stay yours and the signal doesn't survive the rewrite. Paste in a draft, get back clean output, and publish. It's that easy.
Words limit: 0 / 500
Everything you need to catch, strip, and verify ChatGPT's mark, whether it's one page or your whole publishing queue.
One document at a time. Paste in text straight from ChatGPT, get back a structural rewrite with the invisible signal gone and the meaning intact. No account setup required to try it, and no limit on how many times you run a piece through before you're happy with it.
No guessing whether a piece is actually clean. Every rewrite comes back with a detection score attached, checked against the leading AI detectors, so you know before you hit publish instead of finding out after.
Generating through the OpenAI API with watermarking switched on, or working with an EU team whose ChatGPT output is marked by default? Route that content through the Watermark Remover API before it hits your CMS or publish button. Built for teams generating faster than a human editor can rewrite by hand, with turnaround fast enough to sit inside a real production pipeline.
Wait, what's this?
ChatGPT's watermark isn't a logo in the corner of an image. OpenAI calls it textGrain: a secret key influences which words ChatGPT picks as it writes, and hundreds of those decisions add up to a statistical pattern a detector can read. There are no hidden characters or odd spaces to delete. The mark lives in the words themselves, so it travels with the text through copy and paste. OpenAI is rolling it out to ChatGPT and Codex on every plan in the EU, and any API customer, anywhere, can switch it on.
OpenAI has also published where the signal breaks. In its own tests, swapping 10% of the words for synonyms dropped detection from about 92% to 66%, and short passages, math answers, translated text, and heavy edits all weaken it further. That means a light synonym pass still leaves the mark detectable most of the time. Our Watermark Remover doesn't just swap words. It rewrites sentence structure the way a real edit would, which is the specific kind of change textGrain isn't built to survive.
OpenAI's own numbers make the case: swap 10% of the words and the mark is still caught about two-thirds of the time.
Rewrite depth
Invisible-character cleanup
Detection score
Volume
Update cadence
Output readability
Still have questions?
Learn more in our support page.
ChatGPT's watermark is an invisible, machine-readable signal OpenAI embeds in the text ChatGPT and Codex generate, called textGrain. OpenAI introduced it in October 2026 to meet the EU AI Act's transparency rules. It lets OpenAI's detector flag a passage as generated by an OpenAI model, even after it's been copied and pasted somewhere else.
As ChatGPT writes, it chooses each next word from a range of plausible options. textGrain uses a secret key to subtly steer those choices, and across hundreds of words the steering forms a statistical pattern a detector can spot using only the text and the key. Nothing visible is added: no hidden characters, no invisible spaces, no unusual punctuation. The text reads exactly like normal ChatGPT output.
Light editing usually does not remove the watermark. OpenAI built textGrain to hold up through copy and paste and small edits, and in its own tests a 10% synonym swap still left the mark detectable about two-thirds of the time. The mark is broken by the kind of change OpenAI itself flags: substantial rewriting at the sentence level (such as using an AI humanizer like StealthGPT Super), translation, or cutting a passage down until there isn't enough signal left. That's why the Watermark Remover rewrites structure rather than swapping words.
Yes. StealthGPT's Watermark Remover rewrites ChatGPT output at the sentence and clause level, which is the category of change textGrain isn't built to survive, while keeping your meaning and voice intact. Every rewrite comes back with a detection score so you can confirm it's clean before you publish, and teams can run the same process at volume through the StealthGPT API.
One person with one document, a team cleaning up internal documentation, or a whole pipeline running on the OpenAI API, the mark comes out the same way. Enterprise options are available for teams that need it running quietly in the background.
