Does ChatGPT Watermark Its Text? OpenAI’s textGrain Explained

Yes, in the EU. On 5 October 2026 OpenAI announced that it will add an invisible watermark, called textGrain, to ChatGPT and Codex text in the European Union, rolling out over the following weeks, and API developers can switch it on. The mark lives in the model's word choices, not in hidden characters, and OpenAI's own tests show that it fades quickly as the wording changes.

This article covers where the watermark applies, how textGrain works, what OpenAI's numbers say about editing, and what it means if you write with ChatGPT.

What OpenAI announced

OpenAI's announcement, as reported by TechCrunch and BleepingComputer, has three parts:

• ChatGPT and Codex in the EU. OpenAI will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks, for users on all plans.
• API opt-in worldwide. Developers anywhere can turn watermarking on for supported models. It is off by default.
• Limited detector access. OpenAI opened applications for its detector, but access is initially limited to approved researchers and expert organisations.

The reason is the EU AI Act. Its transparency rules took effect on 2 August 2026 and require AI-generated content to be marked so that other systems can recognise it. About two months earlier, Anthropic had already started watermarking Claude's text, worldwide rather than only in the EU.

OpenAI has had a text watermark for a while. The Wall Street Journal reported in 2024 that the company had built one but held back, partly over fears that users would move to competitors that did not watermark.

Where the ChatGPT watermark applies

"Eligible" output is OpenAI's word. It has not listed exactly which responses count, and very short answers are hard to mark in a detectable way anyway.

How textGrain works

textGrain is a statistical watermark. OpenAI published a technical report describing it; in short:

1. While the model writes, a secret key sorts its candidate next words.
2. The model is nudged slightly toward some of those candidates. Readers can't see the difference, and OpenAI says it saw no meaningful change in model quality.
3. A detector that holds the key looks at a passage and adds up hundreds of these small nudges. If the pattern is far stronger than chance, it reports a watermark.

The detector needs only the text and the key. It doesn't need to know who wrote the text or which conversation it came from, and OpenAI says the watermark does not identify the user.

Because the mark is in the choice of words, it survives copy and paste. That also means it does not depend on hidden characters, invisible spaces or unusual punctuation, which is why "watermark removers" that strip those characters do nothing to it.

What OpenAI's tests say about editing

This is the part most people are looking for, and OpenAI was unusually open about it. Its figures, at a 1% false-positive rate:

Short passages, math answers and translated text were all harder to detect, even before any editing.

Two things stand out. First, length matters: a short reply carries fewer nudges, so there is less pattern to find. Second, the mark fades fast. Replacing a quarter of the words, with simple synonym swaps rather than real rewriting, already brought detection below one in five.

That makes sense given the design. Every replaced word removes one nudge and adds a word the key never influenced. Rewrite every sentence in new words and a new structure, which is what a humanizer does, and there is very little of the original pattern left to find. Still, nobody outside OpenAI's approved testers can run the detector, so no tool can honestly promise the mark is gone.

What a watermark proves, and what it doesn't

OpenAI drew the limits clearly:

• A detected watermark shows that an OpenAI system generated or processed part of a passage. It does not say how much of it, or how much human work went into the rest.
• A missing watermark "does not prove human authorship". The text may be too short, edited, translated, or produced by another company's model.

So if someone tells you a ChatGPT watermark was found in your work, it is worth asking how. The detector is not public. What most teachers and editors use is an ordinary AI detector, which estimates from writing style and makes mistakes, including on text people wrote themselves.

What this means if you write with ChatGPT

• Check your rules first. Schools, publishers and employers set their own rules on AI help and disclosure. The watermark does not change them, and rewriting does not either.
• In the EU, assume new ChatGPT text is marked. Once the rollout reaches you, longer ChatGPT and Codex replies will carry textGrain.
• The style gives text away too. Readers and AI detectors notice how ChatGPT text sounds: even sentence lengths, stock transitions, hedging and tidy summaries. Our guide on making ChatGPT writing sound human covers both prompting and editing.
• Character cleaners don't help. textGrain is not made of characters.

If you want the rewrite done for you, the ChatGPT Watermark Remover runs our humanizer on ChatGPT text: it rewrites every sentence in new words, then checks the result against your draft and flags any number, name or citation that changed. You can try it on the page without an account.

The short version

OpenAI is adding an invisible watermark, textGrain, to ChatGPT and Codex text in the EU from October 2026, and API developers can opt in. It lives in the word choices, so stripping characters does nothing. OpenAI's own tests show it fades quickly with editing: a quarter of the words swapped for synonyms brought detection down to about 17%. The detector is not public.