OpenAI said it will begin adding an invisible watermark to text generated by ChatGPT and Codex inside the European Union, in order to comply with the AI Act's transparency obligations, which took effect on August 2 and require AI companies to mark generated content in a way other systems can identify.
The roll-out arrives "over the coming weeks" for eligible ChatGPT and Codex users on all plans — but only in the EU. Developers using OpenAI's API anywhere in the world can switch the watermark on for select models immediately; it is off by default, and the company is deliberately not making text watermarking a global default at launch.
The mark is not a symbol stamped into the output. It works by subtly shaping the model's word choices, nudging next-word predictions with a secret key so that the finished passage carries a statistical pattern no human reader can see but a matching detector can. Because it lives in the words themselves, it travels with the text when it is copied and pasted. OpenAI says the watermark does not identify the user and that it measured no meaningful change in model performance with it enabled.
Alongside the announcement, OpenAI published a technical report for the method, called textGrain, written with researchers from the University of Pennsylvania and Yale. It describes using a key to sort next-word predictions, then adding hundreds of such nudges together until the detector can flag AI-generated content from the text and the key alone.
OpenAI is candid about the limits. In its own tests, replacing 10 percent of words with synonyms dropped detection from about 92 percent to 66 percent. Short passages, mathematical answers and translated text are harder to detect still. Those limitations are the reason initial detector access is being granted only to approved researchers and expert organisations, the company said, so reliability and responsible use can be assessed first. It also cautioned that a missing watermark "does not prove human authorship" — the text could simply be too short, too heavily edited, or produced by a rival's model.
The move lands about two months after Anthropic began watermarking Claude's output worldwide, a decision that drew pushback from users who argued the instructions, context and decisions were theirs and the model was only the tool. OpenAI had built a text watermark before but held back, partly over fears that users would defect to rivals that did not mark their output, according to earlier reporting. Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that have signed up to the EU's code of practice on AI-generated content.
