By Jordan Vale
If you write a work email, a homework answer, or a social post with ChatGPT in the European Union, the words may soon carry a mark you cannot see. OpenAI says it is adding an invisible watermark, called textGrain, to meet the European Union Artificial Intelligence Act (EU AI Act) rules on showing where text came from.
What starts when. From October 5, businesses and developers worldwide that build on OpenAI’s models through its API (the connection that lets other apps use OpenAI’s technology) can choose to turn watermarking on for select models. Over the coming weeks, OpenAI will also add the mark to eligible ChatGPT and Codex (its coding assistant) text generated in the EU. Everywhere else, it stays off by default and is not a global default at launch.
How it works, in plain words. textGrain gently steers which words the model picks, so a detector can later find a statistical pattern in the finished text. It is not a stamp, a label, or a code printed on the page. Think of it more like a handwriting habit: invisible to a reader, but measurable if you have the right tool. OpenAI’s technical report has the math; the everyday point is tracing where text came from, not branding every sentence.
Who can check — for now. At first, only approved researchers and expert organisations get access to the detector. That is on purpose: short passages and later edits make the signal less reliable, so a detector open to everyone could mislead as often as it helps. OpenAI also says math and translation can weaken detection.
What the mark cannot tell you. A detected watermark can suggest that an OpenAI system wrote or processed part of a text. It cannot show who used the tool, how much a person wrote, who owns the text, or whether the content is true. And if no watermark shows up, that does not prove a human wrote it.
Ink, not pencil: this is OpenAI’s stated plan for meeting EU rules on tracing text — not a promise that every AI sentence in Europe will be caught, or that unmarked text is “human.” Treat the mark as one limited signal, with the company’s own caveats attached.
Why regular people should care
The quiet assumption is changing. Students, freelancers, and office workers in the EU already draft with ChatGPT, often assuming AI text leaves no trail. A hidden mark changes that. It matters for teachers, employers, and online platforms trying to tell machine-made writing from human writing — as long as no one pretends the line is sharp. And a missing mark is not proof that a person wrote something.
What's next
What to watch. How fast EU ChatGPT and Codex text actually gets the mark, who gets detector access after the first group, and how other AI companies answer the same EU AI Act rules. The facts today are narrow: opt-in for developers now; EU ChatGPT and Codex text over the coming weeks; off by default elsewhere — with clear limits on what the mark can and cannot prove.