OpenAI will start watermarking text from ChatGPT and Codex for users in the EU in the next few weeks, to meet transparency rules under the EU AI Act that took effect August 2. The method, called textGrain, nudges how the model picks words into a pattern that a detector holding a secret key can find, and it survives copy and paste. Developers anywhere can turn it on in OpenAI’s API for select models, but the setting starts switched off. OpenAI’s own tests show the mark is easy to weaken, with a 10% synonym swap cutting the detection rate to 66% from 92% and a 25% swap dropping it to 17%, so only approved researchers get the detector initially. Anthropic is applying its Claude watermark worldwide, but OpenAI isn’t making the feature a global default.
