Invisible watermark coming to ChatGPT texts in the European Union

Phone callbackI call you at the time you choose
← Back to the blog
Blog

Invisible watermark coming to ChatGPT texts in the European Union

Author: Lakatos Szilárd | Published:

Soon every text written with ChatGPT in the European Union will carry a mark invisible to the naked eye. That's how OpenAI is complying with the transparency rule of the EU AI Act – but the solution is far from foolproof.

According to the company's announcement, in the coming weeks texts produced by ChatGPT and Codex in the EU will automatically be watermarked, on every subscription tier from free to paid. Developers can already turn the feature on worldwide via the API, but there it's off by default.

Why now?

Behind it is the EU AI Act. The transparency rules that took effect on 2 August require AI developers to label the content they generate in a machine-detectable way. For images and video there are already established solutions, but for text a workable method had been missing.

How do you mark a text invisibly?

The watermark isn't a hidden character or metadata – it hides in the wording itself. The textGrain method uses a secret key to subtly influence which of the possible next words the model chooses. The reader notices nothing and the text reads naturally, but a purpose-built detector can identify the pattern hidden in the word choice.

Because the mark lives in the order and choice of words, it survives copying and pasting. OpenAI researchers developed the method together with experts from the University of Pennsylvania and Yale. The company stresses that the watermark doesn't identify individual users; it only signals that the text came from its tools.

How reliable is it?

OpenAI itself admits the system has serious limits. According to its own measurements:

  • on untouched text, detection is about 92 percent;
  • if a tenth of the words are replaced with synonyms, it falls to 66 percent;
  • after replacing a quarter of the words, only 17 percent remains.

The watermark is also hard to detect in short texts, translated content and maths-style answers. That's why the company explicitly warns: the absence of a watermark doesn't prove that a human wrote the text.

Who can check it?

For now, almost no one. Only approved researchers and professional organisations can access the detector, on application, for evaluation purposes. So there's no public checking tool: no teacher, editor or client will be able to run the test themselves.

What does this mean in practice?

For anyone using ChatGPT to write letters, brainstorm or draft, nothing changes: answers look just as before. The watermark doesn't forbid using the text, and no label appears in it.

But anyone passing off machine-written text unchanged as their own – on a website, in a paper, in a proposal – should know that its origin may become detectable over time. For websites this is a particularly interesting question, since search engines and clients pay ever more attention to whether real expertise stands behind the content.

The lesson is the same as ever: artificial intelligence is a good assistant, but the final text is best written in your own voice, from your own experience. If you'd like help with your website's copy, feel free to get in touch.

chatgptartificial intelligenceAIAIopenaiEuropean Unionprivacyregulationwatermarkwebsite design

More IT news and tips on the blog

More articles

See all articles →