OpenAI is adding an invisible watermark to ChatGPT and Codex text output within the European Union to meet new EU AI Act requirements.
The rollout
Transparency rules introduced on August 2 mandate that AI firms mark their content so other systems can spot it. OpenAI plans to deploy this feature over the coming weeks for eligible users on all plans, but the change applies only to the EU region.
Developers using the API can enable the feature for specific models starting today. It is currently switched off by default, and OpenAI is not making it a global standard at this stage.
How it works
The watermark is not a visible symbol. Instead, it subtly alters the model’s word choices to create a pattern invisible to the reader but detectable by software. Because the mark exists within the words themselves, it travels with the text when copied and pasted.
OpenAI confirmed the process does not identify individual users. The company stated that enabling the watermark caused no meaningful drop in model performance.
A technical report accompanies the announcement, detailing a method called textGrain. Developed with researchers from the University of Pennsylvania and Yale, the system uses a secret key to influence next-word predictions. By applying hundreds of these small nudges, a detector can identify AI-generated content using only the text and the key.
Limitations and detection
Editing the text can remove the watermark. Tests show that replacing 10% of words with synonyms drops detection rates from roughly 92% to 66%.
Short passages, mathematical answers, and translated text also prove harder to detect. Consequently, OpenAI is limiting initial detector access to approved researchers and expert organisations to evaluate reliability and responsible use.
The company warns that a missing watermark does not prove human authorship. Text may be too short, heavily edited, or generated by another AI provider. A watermark can indicate that an OpenAI system processed a passage, but it cannot quantify the level of human judgment or creativity applied.
Industry context
This move follows Anthropic’s decision two months ago to watermark Claude text globally. That decision faced backlash from some users who argued they provided the instructions and context while the tool simply executed the task.
OpenAI previously built a text watermark but delayed its release. The Wall Street Journal reported in 2024 that the company feared users might switch to rivals offering unwatermarked alternatives.
Anthropic, Google, Meta, Microsoft, and OpenAI have all committed to following the EU’s code of practice on AI-generated content.
