OpenAI is adding an invisible watermark to text from ChatGPT and Codex in the European Union, so that other systems can tell the text was written by AI. It says it is doing this to follow the EU AI Act.
Major Highlights
- Why: The EU AI Act’s transparency rules, in force since August 2, require AI companies to mark AI-generated content in a way other systems can identify.
- Who gets it: The watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans, in the EU only.
- Developers: Developers using OpenAI’s API anywhere in the world can switch it on for select models starting now. It is off by default, and OpenAI says it is not making watermarking a global default at launch.
- How it works: You can’t see it. It quietly shapes the model’s word choices into a pattern that only a detector can pick up, so it stays with the text when you copy and paste. OpenAI says it doesn’t identify the user and didn’t noticeably change the models’ performance. The method is called textGrain, and OpenAI published a technical report on it with researchers from the University of Pennsylvania and Yale.
- It can be weakened: In OpenAI’s own test, replacing 10% of the words with synonyms dropped detection from about 92% to 66%. Short passages, maths answers and translated text are also harder to detect.
- Who can check: For now, only approved researchers and expert organisations get access to the detector.
- Missing watermark means nothing: OpenAI says no watermark does not prove a human wrote the text. It could be too short, heavily edited or from another company’s AI.
- Anthropic already did it: Two months ago, Anthropic said it would watermark Claude’s text worldwide. Some Claude users pushed back.
KINI BIG DEAL
This is the first real sign of how AI rules in one region reach everyone else. The EU wrote the rule, and OpenAI is applying it only where the law makes it. Anthropic chose to apply its version worldwide, while OpenAI is keeping its default to the EU for now.
For us, that means a ChatGPT user in Nigeria is not watermarked for now, while an API developer here can switch it on. If you work in a newsroom, school or company that worries about AI-written text, don’t expect a simple way to catch it yet. The detector isn’t public, and a light edit can weaken the mark.
