OpenAI will start hiding a watermark in text written by ChatGPT and Codex, but only for people in the European Union. The company laid out the plan on October 5th: the mark rolls out “over the coming weeks” to eligible users on every plan in the EU, in response to the EU AI Act.
The system is called textGrain, and you can’t see it. It nudges which words the model picks, leaving a statistical pattern only OpenAI’s detector can spot. Because it lives in the wording, it survives copy and paste. Short replies get a pass: the EU’s code of practice exempts code snippets and anything under 200 tokens (the word pieces AI models read and write), about 150 words in English.
Outside Europe, only devs get a say: API customers can now switch watermarking on for select models, but it’s off by default. OpenAI said it isn’t making it “a global default at launch,” and that the regional approach “gives us room to learn from real-world use and feedback.” The AI Act drew the map, and it leaves out all of Asia, including South Korea, whose AI Basic Act took effect January 22nd and calls for watermarks or similar labels on AI-generated content, according to AFP. Anthropic went the other way in August, watermarking Claude’s text worldwide “because we don’t yet have a durable way to scope it by region.”
The take
None of this is news to OpenAI. In August 2024, it said its text watermark could be dodged with a translator or a rewrite by another AI, and that it could “stigmatize use of AI as a useful writing tool for non-native English speakers.” By their own charts, the limits haven’t gone anywhere. It’s just that the law has changed. We’ll give them credit for being upfront about the gaps. But if you were hoping for a reliable ChatGPT detector, sadly, this is still not it.
