OpenAI adds textGrain, an invisible watermark, to the text generated by ChatGPT and Codex in the European Union. The rollout will be automatic in the coming weeks and without an option to deactivate it.
OpenAI has announced that it will add textGrain, an invisible watermark, to the text generated by ChatGPT and Codex in the European Union. The deployment, as reported by ecosistemastartup.com, will arrive in the coming weeks and will cover all plans within the community territory.
The watermark is created during generation by influencing word choice through a secret key and prior context. No hidden characters or invisible spaces are inserted, so cleaning the text does not remove it. Only a detector with the key can find it, and for now, it is not public: OpenAI limits access to approved researchers and organisations.
The deployment has three levels. In ChatGPT and Codex within the EU, it will be automatic and without an option to deactivate it. In the API, it is optional for clients worldwide, turned off by default and only for selected models from October 5. Outside the EU, this will not be the default behaviour, at least at launch.
OpenAI plans to release textGrain as open source, although no date has been set.
The watermark's resistance is limited. With a false positive rate of 1%, the detector identified nearly 80% of 200-token passages and around 95% of 400-token passages in psychology content. Detection was worse in mathematics because there are fewer ways to say the same thing.
Editing weakens it quickly. In 400-token passages, changing 10% of the words to synonyms dropped detection from 92% to 66%; changing 25% left it around 17%. Translating, shortening, or rewriting also reduces the signal.
"A watermark that disappears with a rewrite protects against careless use, not against malicious use"
The reason for applying it only in the EU is Article 50(2) of the European AI Regulation, which requires marking synthetic content in machine-readable format. These obligations apply from August 2, 2026. The law does not require a specific technology, but the European Commission developed a voluntary Code of Practice on Transparency of AI-Generated Content, signed by Anthropic, Google, Meta, Microsoft, Mistral, and OpenAI.
Systems prior to August 2 have a grace period until December 2, 2026, introduced by the Digital Omnibus package. OpenAI's upcoming weeks fall within that window. Fines for non-compliance can reach 15 million euros or 3% of global turnover.
OpenAI is late compared to its direct rival: Anthropic applies invisible watermarking to all texts generated by Claude and implements it globally, not just in the EU.
For a small or medium-sized enterprise or startup publishing content, the change is manageable. If you generate bulk text via API, the watermark is not included unless you activate it. If your team works both inside and outside the EU, you will have marked texts and others that are not. The watermark can disappear with a rewrite, but the absence of a watermark does not prove that the text was written by a person.
You may also like
