OpenAI Introduces Invisible Text Watermarking for ChatGPT in the EU to Comply with AI Act

Image: TechCrunch · Source

OpenAI will embed invisible watermarks in ChatGPT and Codex outputs within the EU, enabling AI content identification per new regulations, while allowing selective detection and acknowledging limits when edited or shortened.

OpenAI is implementing an invisible watermarking technology for text generated by its ChatGPT and Codex models within the European Union to comply with the EU AI Act's transparency requirements that took effect in August 2026. This legislative measure mandates AI companies to label AI-generated content in a way that other systems can detect.

The watermarking will be gradually introduced to eligible ChatGPT and Codex users across all subscription levels but exclusively in the EU region. Meanwhile, developers utilizing OpenAI’s API worldwide can activate watermarking on select models starting immediately, though it remains off by default globally. OpenAI has no plans to enforce watermarking universally at this stage.

Unlike typical visible marks, the watermark is embedded by subtly influencing the AI’s word choice patterns. This creates an imperceptible signature present in the text itself, allowing detection tools to identify AI-generated passages even after copying or pasting. OpenAI emphasizes the watermark does not track user identities and does not noticeably affect model output quality.

To support transparency, OpenAI released a technical report detailing their watermarking approach, named textGrain. Developed jointly with researchers from the University of Pennsylvania and Yale, the method applies a secret key to reorder next-word probabilities, embedding patterns throughout the text. Detection relies solely on the text and this secret key.

However, OpenAI acknowledges limitations: editing can reduce watermark detectability. For example, replacing 10% of words with synonyms dropped detection accuracy from about 92% to 66%. Additionally, shorter text segments, mathematical answers, and translated content pose detection challenges. Due to these factors, initial access to watermark detection is restricted to approved researchers and expert entities to ensure responsible evaluation.

OpenAI stresses that absence of a watermark does not confirm human authorship since content could be heavily edited, too brief, or generated by other AI providers. The watermark only signals that an OpenAI system was involved in producing or processing the text, not the extent of human input or creativity.

This move follows a similar step by Anthropic, which began watermarking the AI model Claude worldwide in August 2026, although that initiative met criticism from users emphasizing their role in providing instructions and context. OpenAI previously developed watermarking but delayed rollout partly due to concerns about users migrating to non-watermarking competitors.

Industry leaders including Anthropic, Google, Meta, Microsoft, and OpenAI have pledged adherence to the EU’s code of practice for AI-generated content, aligning with increasing demands for AI transparency and accountability.