What is textGrain? OpenAI's text watermark, explained
A simple guide to the hidden watermark OpenAI announced on 5 October 2026. How it works, where it is turned on, how well it works, and what it cannot tell anyone.
textGrain is a hidden watermark that OpenAI adds to AI-written text. It changes which words the model picks, in a pattern made with a secret key. OpenAI’s detector can spot that pattern. It is on for ChatGPT and Codex text in the EU, and API customers anywhere can turn it on.
What textGrain is
textGrain is OpenAI’s name for a watermark in text. A watermark on a photo is something you can see. This one you cannot see, select or delete. It is in the choices the model makes while it writes.
OpenAI announced it on 5 October 2026. It is coming to ChatGPT and Codex text for users in the European Union over the following weeks. API customers anywhere can also turn it on for some models.
How it works, without the maths
An AI model writes one small piece of text at a time. At each step there are often several good options, like “quick”, “brief” or “short”. textGrain uses a secret key to give some options a tiny push. One push means nothing. Hundreds of them across a text add up to a pattern that is very unlikely to happen by chance.
OpenAI’s detector runs the same key over a text and measures how strong the pattern is. It needs the text and the key. It does not need the original chat.
Want to see the idea in action? Our watermark demo does the same thing with a public key and a short word list.
Who gets it
| Product | Watermarked? |
|---|---|
| ChatGPT, users in the EU | Yes, from October 2026, on all plans |
| Codex, users in the EU | Yes, from October 2026 |
| ChatGPT and Codex outside the EU | Not by default at launch |
| OpenAI API | Only if you turn it on, for some models, anywhere |
OpenAI has not shared the full list of API models. It has also not said if EU ChatGPT users can turn the watermark off.
How well it works
OpenAI set its detector so that only about 1% of text without a watermark gets flagged. At that setting, OpenAI reported:
- About 80% of 200-token texts found, and about 95% of 400-token texts, on normal topics like psychology.
- Much lower results for maths, where there is less choice of words.
- On 400-token texts: about 92% found with no changes, 66% after 10% of the words were changed to synonyms, and 17% after 25% were changed.
OpenAI also said the quality of answers stayed about the same with the watermark on, across eight tests. Our detectability calculator turns these numbers into a chart you can play with.
What it cannot tell anyone
OpenAI is very clear about the limits. A textGrain result does not:
- show the user, account, prompt or chat;
- show how much a person wrote or changed;
- show who owns the text or who is responsible for it;
- show if the text is true.
It also works the other way. A text without a watermark does not prove a person wrote it. It may be short, edited, translated, from another AI, or from ChatGPT outside the EU.
Who can use the detector
Right now, nobody can paste text into a public textGrain checker. OpenAI is taking applications for its detector, but only approved researchers and expert groups get access, one by one. OpenAI says this is because the detector can make mistakes both ways.
So if a website says it can detect “the ChatGPT watermark”, it is guessing from writing style or looking at hidden characters. Neither of those is textGrain.
Why now: the EU AI Act
Article 50(2) of the EU AI Act says companies that make AI tools must mark AI output so it can be detected. This applies from 2 August 2026. Tools that were already on sale have until 2 December 2026. textGrain is how OpenAI follows this rule for text. Google (SynthID) and Anthropic (Claude) have their own ways. See textGrain vs SynthID.
In 2024, OpenAI said it had built a working text watermark but did not use it. It was worried about things like unfair flags for people who write English as a second language, and users moving to other apps. The EU rules changed that.
What about hidden characters?
Hidden characters like zero-width spaces do show up in AI text. But they are not textGrain. They come from formatting and copying, and you can remove them safely with our hidden character remover. Removing them does not touch textGrain, which is in the words.
Questions people ask
Can I see textGrain in my text?
No. There is nothing to see: no characters, no formatting and no file data. The text looks exactly like normal ChatGPT text.
Can I turn textGrain off?
OpenAI has not said if EU ChatGPT users can turn it off. In the API it is off unless you turn it on.
Can my teacher check my essay for textGrain?
Not directly. For now, only approved researchers and expert groups can use the detector. Be careful with any tool that says it can check it.
Does textGrain work in other languages?
OpenAI has not shared results for each language. The method only changes word choices, so it should work in any language. But detection rates may be different.
Is textGrain open source?
Not yet. OpenAI published a technical report and says it plans to share the technology as open source. It has not given a date.