OpenAI to watermark ChatGPT and Codex text in EU

OpenAI's decision to watermark text from ChatGPT and Codex is a direct response to the EU AI Act. The company will roll the feature out to all plan users in the EU within weeks, while developers using OpenAI's API can enable it on some models—though it remains off by default. OpenAI has said it has no plans to make text watermarking default worldwide at this time.

The watermark works by subtly altering the model's word choices to embed a pattern invisible to human readers but detectable by tools. Because it is built into the words themselves, it survives copy-pasting. OpenAI claims it will not identify users and has little effect on model performance. Alongside the announcement, OpenAI published a report on textGrain, its detector for watermarked text.

However, the method has clear limits. Swapping synonyms sharply reduces detection: unedited text is detected 92% of the time, but replacing 10% of words drops that to 66%, and replacing 25% drops it to 17%. OpenAI also says short texts, math answers, and translated text are harder to detect. Because of these constraints, access to the detector is initially limited to approved researchers and institutions. OpenAI warns that the absence of a watermark does not prove human authorship, and that a watermark shows AI involvement but cannot measure how much human judgment, editing, or creativity went into the final text. How well the system performs in real-world use—and whether it can be easily evaded—will determine its value for compliance and detection.