OpenAI is going to start hiding a watermark inside text written by ChatGPT and Codex in the European Union. TechCrunch reports that the company announced the change Monday in a blog post, and the reason is legal. The EU AI Act's transparency rules took effect on August 2, and they require AI companies to mark AI-generated content in a way other systems can identify.
The rollout happens over the coming weeks for eligible ChatGPT and Codex users on all plans, but only in the EU. Developers who use OpenAI's API anywhere in the world can switch it on for select models starting now. It is off by default there, and OpenAI said it is not making text watermarking a global default at launch.
So what is a text watermark? It is not a symbol or a hidden character. OpenAI's method works by gently steering which words the model picks. You will not notice anything when you read it, but a detector that knows the pattern can. Because the signal lives in the words themselves, it survives a copy and paste. OpenAI said the watermark does not identify the user, and that it saw no meaningful change in how well its models performed with it turned on.
The company also published a technical report on the method, which it calls textGrain. It was co-written with researchers from the University of Pennsylvania and Yale. The idea is that a secret key is used to sort the model's next-word options as it finishes a sentence. One nudge means nothing. Hundreds of them add up to a pattern the detector can spot using only the text and the key.
Here is the part I appreciate. OpenAI was upfront about where this breaks. In its own tests, swapping 10% of the words for synonyms dropped detection from about 92% to 66%. Short passages, math answers, and translated text are harder to catch. Because of those limits, OpenAI is giving detector access only to approved researchers and expert organizations for now.
It also warned against reading too much into a clean result. A missing watermark "does not prove human authorship," the company said. The text could be too short, too heavily edited, or written by another company's AI. And a watermark that is present only tells you an OpenAI system touched part of a passage. In OpenAI's words, it does not show "how much human judgment, editing, or creativity went into it."
OpenAI is not first here. Two months ago, Anthropic said it would watermark text from Claude, and it is applying that worldwide. Some Claude users pushed back, arguing they supplied "the instructions, context, decisions" while Claude was just "the tool." OpenAI had built a text watermark before and held back, partly over worry that users would leave for rivals that did not watermark, according to a 2024 Wall Street Journal report. Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have committed to the EU's code of practice on AI-generated content.
As a developer, I think this is a sensible first step, mostly because nobody is overselling it. A watermark is a hint, not a verdict. Treat it like a smoke detector. It is useful when it goes off, and silence does not mean there is no fire.
If you live in the EU and use ChatGPT for school or work, nothing about how you use it changes. Your writing still belongs to you, and the watermark is not tied to your name. If you are a teacher or an editor, the honest takeaway is that there is still no magic test for AI text, so keep judging the work on its merits. And if you build on the API, you now get a real choice to label what your app produces. I would flip it on for anything public facing. Being clear about what a machine wrote is a cheap way to earn trust.
