OpenAI starts watermarking ChatGPT text in the EU

OpenAI has announced that it will begin adding invisible watermarks to text generated by ChatGPT and Codex within the European Union. This move is designed to ensure compliance with the strict transparency mandates outlined in the EU AI Act, which officially took effect on August 2 and requires artificial intelligence companies to properly mark machine-generated content so that other systems can easily identify it.

The rollout of this new feature is scheduled to happen over the coming weeks, targeting eligible ChatGPT and Codex users across all plans specifically within the EU region. For developers utilizing OpenAI’s API anywhere in the world, the watermarking capability can be manually activated for select models starting immediately, though it remains turned off by default. OpenAI has decided against making text watermarking a global default setting during this initial launch phase.

Rather than appearing as an actual visual symbol on the screen, the watermark operates by subtly shaping the model's word choices during generation. This process leaves behind an invisible pattern that human readers cannot see, but specialized detectors are able to pick up. Because the watermark is embedded directly within the words themselves, it successfully travels along with the text even when it is copied and pasted into other applications or documents.

OpenAI has confirmed that the watermark does not track or identify the user, and internal testing showed no meaningful degradation in model performance when the feature was switched on. Alongside this announcement, the company published a comprehensive technical report detailing its watermarking approach, known as textGrain, which relies on entropy-calibrated techniques. Co-authored with researchers from the University of Pennsylvania and Yale, the report explains how a secret key is used to sort next-word predictions, creating a verifiable pattern when hundreds of subtle nudges are combined.

Despite its technical sophistication, the watermark does have notable limitations. OpenAI’s testing suggests that editing the text can significantly reduce detectability. For instance, replacing just 10% of the words with synonyms dropped the detection success rate from approximately 92% to 66%. Additionally, the company noted that short passages, mathematical answers, and translated text are inherently harder for the detectors to accurately identify.

Due to these technical constraints, OpenAI is initially restricting detector access to approved researchers and expert organizations. These groups will assist the company in evaluating the reliability and responsible use of the technology. OpenAI also cautioned that the absence of a watermark does not constitute definitive proof of human authorship, as the text might simply be too short, heavily edited, or generated by a competing AI model.

This development follows a similar announcement made two months prior by rival firm Anthropic, which stated it would implement text watermarking for its Claude model on a global scale. That decision encountered some backlash from users who felt penalized for using AI as a productivity tool. Industry giants including Anthropic, Google, Meta, Microsoft, and OpenAI have all committed to following the EU’s code of practice regarding AI transparency.