AI Text Watermark: OpenAI's textGrain Explained

Your ChatGPT draft may soon carry baggage you can't see, delete, or explain away with "but I edited it."

OpenAI is adding a hidden statistical watermark to text produced for users in the European Union. Not everywhere. Not instantly. But if AI touches your scripts, newsletters, captions, client copy, or code, this one deserves attention.

The quiet mark

OpenAI announced textGrain on October 5¹. The rollout will cover eligible ChatGPT and Codex output across all plans in the EU over the coming weeks. API customers worldwide can switch it on, although it remains off by default.

This isn't a sneaky character hiding between two words. textGrain slightly steers the words or word fragments the model selects, creating a pattern that OpenAI's detector can recognize. Copying, pasting, or changing the font won't scrub it.

The move follows the EU AI Act's transparency rules, which started applying on August 2, 2026. The EU's voluntary compliance code² has roughly 190 signatories and covers machine-readable marking plus labels for certain AI-generated public-interest content.

OpenAI isn't first. Google has used SynthID for AI text, while Anthropic said new Claude models will carry watermarks³. Claude's wider approach already annoyed writers and developers who worry that editing their own work with AI could leave the whole piece looking machine-made.

Your byline, their signal

Here's the awkward bit: the watermark can show that OpenAI processed some text. It cannot show how much you wrote, whether you only fixed grammar, or whether you spent six hours rebuilding the draft afterward.

It also doesn't expose your identity, account, prompt, or conversation. And it doesn't prove the text is accurate. A watermark says "AI was probably involved." That's it.

Detection is shaky, too. In OpenAI's tests, roughly 92% of longer watermarked passages were caught. Replacing 10% of the words with synonyms dropped that to 66%. Replacing 25% pushed it down to 17%. Short pieces are harder to judge.

Don't panic. But stop treating paste-and-publish as a creative workflow. Frankly, you should've stopped already.

The detector won't be public at launch. OpenAI is limiting access to approved researchers and specialist organizations because false positives and missed marks remain possible. So expect confusion before you expect certainty.

Do this now

Keep your drafts and revision history. If a client or platform questions your work, a trail showing your research, notes, recordings, and edits beats an argument about detector scores.

Use AI for options, structure, and critique - not as the final keyboard. Rewrite from your own thinking instead of polishing a finished AI draft until it sounds vaguely like you.

Finally, check client contracts and platform rules. Ask what "AI-generated" actually means to them. Assistance and authorship aren't the same thing, even if today's watermark can't tell the difference.