If you use ChatGPT in Europe, the text it writes for you is about to carry a hidden signature. OpenAI confirmed on October 5 that a ChatGPT watermark called textGrain will roll out to EU users over the coming weeks. You won’t see it, it won’t change how answers read, and it’s far easier to break than the headlines suggest.
Here’s what you need to know in ten seconds:
- Who: EU users on every ChatGPT and Codex plan, free and paid.
- When: Rolling out over the coming weeks, ahead of the EU AI Act’s December 2, 2026 compliance date for existing AI providers.
- Outside the EU: No default watermark in ChatGPT. API developers anywhere can opt in.
- Can you check text? Not yet. The detector is limited to approved researchers at launch.
- Easy to strip? Yes. Changing a quarter of the words drops detection to about 17%.
How the ChatGPT Watermark Actually Works
Forget a visible label or a tag hidden in the file. textGrain works inside the writing itself. When ChatGPT picks its next word, there are usually several good options, and the watermark gently nudges the model toward certain ones.
Any single nudge is meaningless. But as OpenAI put it, per TechCrunch, “add hundreds of these nudges together, and the detector can spot AI-generated content.” The detector only needs the text and a secret key held by OpenAI.
That’s why the signal lives in the statistics, not in any one sentence. You can read a watermarked paragraph a hundred times and never notice a thing.
Why Europe Gets It First
This isn’t OpenAI volunteering. The EU AI Act’s transparency rules (Article 50) require AI-generated content to be marked in a machine-readable way. Those rules started applying on August 2, 2026, and providers already on the market have until December 2, 2026 to comply, according to The Next Web.
OpenAI had reportedly built a text watermark years ago and shelved it. One worry was that users would simply switch to chatbots that didn’t watermark. Now that the law applies to every major provider in the EU, that worry mostly goes away.
OpenAI also isn’t alone. Anthropic said in August it would add an invisible watermark to Claude’s text, and The New Stack reports Anthropic applies it on every Claude product with no opt-out. If you’ve been weighing how AI companies handle your data, our breakdown of the Claude voice data training opt-in covers another side of that picture.
How Reliable Is Watermark Detection?
This is where the story gets more interesting. OpenAI published its own numbers, and they’re honest about the weak spots. All figures below are at a 1% false-positive rate.
| Scenario | Detection rate |
|---|---|
| Untouched 200-token passage (~150 words) | About 80% |
| Untouched 400-token passage (~300 words) | About 92–95% |
| 10% of words swapped for synonyms | About 66% |
| 25% of words swapped | About 17% |
| Math answers, short or translated text | Noticeably weaker |
The takeaway: longer, untouched text is easy to flag. Light editing cuts that sharply, and a solid rewrite all but erases it. Math is hard to mark because there are only so many ways to write an equation.
What the Watermark Can’t Tell Anyone
OpenAI was unusually clear about the limits, and they matter if you’re a student, a writer, or someone who edits AI drafts for work. According to Tech Startups, the watermark cannot:
- Identify you, your account, or who ran the prompt.
- Prove who owns the text or who is legally responsible for it.
- Check whether the content is accurate.
- Measure how much a human edited or added.
One more point is easy to miss. A missing watermark does not prove a human wrote something. Text could come from another tool, be edited, or simply be too short to test.
What Developers Need to Know About the API
If you build on OpenAI’s API, nothing changes unless you choose it. The New Stack reports that watermarking is off by default and can be switched on per project or organization for supported models. Once on, you don’t change your requests at all.
OpenAI says it saw no meaningful hit to coding benchmarks with the watermark on. It also admits code is harder to mark than prose, for the same reason math is. The company plans to open-source textGrain later, built with researchers from the University of Pennsylvania and Yale.
Monetization and transparency are moving at the same time at OpenAI, too. We recently covered the company’s test of ads in ChatGPT image generation, which hits a different set of users.
Frequently Asked Questions
Will the ChatGPT watermark change how my answers read?
No. It only shifts which of several good word choices the model picks. OpenAI says readers can’t see it, and output quality isn’t meant to change.
Is ChatGPT text watermarked in the US or UK?
Not by default at launch. The automatic watermark applies to eligible EU users. API developers worldwide can turn it on themselves.
Can my teacher or boss check if I used ChatGPT?
Not with this tool, at least for now. There’s no public checker, and the detector is limited to approved researchers and expert groups.
Can I turn the watermark off in ChatGPT?
OpenAI hasn’t announced an opt-out for EU ChatGPT users. Because it’s a legal requirement there, don’t expect one.
Does editing remove the watermark?
Mostly, yes. Swapping 10% of words drops detection to about 66%, and 25% drops it to about 17%.
Our Take
For everyday users, this changes almost nothing: the text looks the same, nobody can check it yet, and a normal edit weakens the signal fast. The real shift is that all the big AI labs now mark their writing in Europe, which makes “AI or not?” a question with partial, provable answers instead of guesswork. If you use ChatGPT for work, the smart habit hasn’t changed. Treat its output as a first draft, rewrite it in your own voice, and be upfront when AI helped.
Leave a Reply