OpenAI will start watermarking ChatGPT’s text in the EU

OpenAI will begin embedding an invisible watermark into text produced by ChatGPT and Codex for users in the European Union to comply with the EU AI Act, the company said. The watermark, implemented via a method called textGrain, subtly biases word choices so a detector with a secret key can identify AI-generated passages, though editing and certain content types reduce detectability.

By AI Newsroom· Reviewed by Pranav, Founder & Editor-in-ChiefPublished less than a minute agoUpdated less than a minute ago0 views

OpenAI will start watermarking ChatGPT’s text in the EU

Why It Matters

This change implements the EU AI Act's transparency requirement that AI-generated content be markable by other systems, setting a compliance precedent for major AI providers and affecting how AI outputs can be identified and regulated in the EU. The limits OpenAI reports on detection accuracy underscore challenges for enforcement and for developers deciding whether to adopt watermarking globally.

Key Facts

Regulation prompting change: EU AI Act transparency rules took effect on August 2
Products affected: ChatGPT and Codex text in the European Union
Rollout timing: Watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans in the EU
API option: Developers using OpenAI’s API worldwide can enable the watermark for select models starting today; it is off by default
Global default: OpenAI is not making text watermarking a global default at launch

OpenAI announced it will add an invisible watermark to text generated by ChatGPT and Codex for users in the European Union to meet the EU AI Act’s transparency requirements. The rule, which came into force on August 2, requires AI companies to mark AI-generated content in a machine-detectable way. OpenAI said the watermarking feature will appear for eligible EU users across plans over the coming weeks, while API customers worldwide may enable it for select models starting immediately, though it will be disabled by default.

The watermark is not a visible tag but a pattern embedded in the model’s word selection. OpenAI described the technique in a technical report called textGrain, co-authored with researchers at the University of Pennsylvania and Yale. The method uses a secret key to reorder likely next-word choices and nudges generation toward words that create a detectable statistical pattern; a detector holding the key can then identify text that carries that pattern.

OpenAI cautioned that the watermark is imperfect. In internal tests, substituting roughly 10% of words with synonyms reduced detection rates from about 92% to 66%. The company also noted that short passages, mathematical answers, and translated text are harder for the detector to flag. Because the mark lives in the text itself, it persists when content is copied and pasted, and OpenAI said it does not encode user identity nor noticeably affect model performance when enabled.

OpenAI limited initial detector access to approved researchers and expert organizations to help evaluate reliability and responsible use. The company warned that an absent watermark does not prove human authorship, since heavy editing, short length, or output from a different provider could explain non-detection. The announcement follows Anthropic’s earlier decision to watermark Claude globally, and comes as multiple large AI companies have committed to the EU’s code of practice on AI-generated content.