ChatGPT watermark: OpenAI will sign your text in secret with textGrain

The ChatGPT watermark is an invisible signature OpenAI will add to the text ChatGPT generates, and to Codex text too. It was announced on October 5, 2026, starts in the European Union, applies to all plans and rolls out over the coming weeks. Here is how it works, how well it holds up and who can read it.

How textGrain works

It is called textGrain. According to OpenAI, it uses a secret key to nudge which word comes next, so the pattern is hidden in the words the model picks. It is not a visible stamp or added text: the signature lives inside. That is why it survives copy and paste: moving the text to another document does not erase what it carries inside.

Who can read the watermark

For now, only approved researchers and organizations get the detector. In the API, developers can turn the mark on worldwide, but it is off by default. In other words, even if a text carries the signature, most people have no way to check it.

How well it holds up: it is not foolproof

On unedited 200-token passages, OpenAI reports about 92% detection. If you swap one in ten words for synonyms, it drops to 66%. And, per OpenAI, a missing mark does not prove a text is human-written. The mark says an AI took part, not how much judgment the person who reviewed and edited the text put in.

How to apply it in your business?

If your team writes emails, proposals, posts or documents with help from ChatGPT or Codex, list which texts you make with AI each week and decide for which ones you would care that the text carries a signature. If you work with clients in the European Union, review your editing process: what counts is the human judgment that edits and approves, not whether the text carries a mark. For another case of how AI-made content gets verified, read the full breakdown of the iPhone 18 Pro's photo verification.

Frequently asked questions

What is textGrain?

It is the method OpenAI uses to hide an invisible signature in ChatGPT and Codex text, using a secret key to nudge which word comes next.

Where does the ChatGPT watermark start?

In the European Union, on all plans, and it rolls out over the coming weeks.

Does it survive copy and paste?

Yes, according to OpenAI the mark survives copy and paste.

Who can detect it?

For now, only approved researchers and organizations. In the API, developers can turn it on worldwide, off by default.

Conclusion

The ChatGPT watermark is an invisible signature that is not foolproof: change one in ten words and detection drops from 92% to 66%. It says an AI took part, not how much judgment you put in. Tell me in the comments what texts you write with AI each week and I'll tell you if the mark matters to you.