OpenAI will add watermarks to text generated by ChatGPT and Codex in the EU

OpenAI announced it will add invisible watermarks to text generated by ChatGPT and Codex in the EU, to comply with the EU AI Act transparency rules that take effect on August 2. The watermark will roll out to eligible users across all plans in the EU in the coming weeks. Global API developers can enable it for some models starting today, with it off by default, and it will not be set as a global default for now. The watermark works by fine-tuning the model's word choices to form patterns recognizable by a detector, and it persists when text is copied and pasted. OpenAI says it does not identify users and that no significant change in model performance has been observed. OpenAI also published a technical report, textGrain, co-authored with researchers from the University of Pennsylvania and Yale, describing a method that uses a key to rank next-token predictions. Tests show that replacing 10% of words with synonyms reduces detection from about 92% to 66%, and that short texts, mathematical answers, and translated texts are harder to detect, so the detector will initially be available only to approved researchers and expert institutions.