OpenAI is set to introduce an imperceptible watermarking system for the content produced by its AI models, ChatGPT and Codex, within the European Union. This initiative is a direct response to the EU AI Act's mandate for greater transparency regarding AI-generated material. While the watermarks are designed to be undetectable by human readers, their effectiveness can be diminished through significant textual alterations.
OpenAI Introduces Invisible Watermarking for AI-Generated Text in the EU
In a significant development on October 5, 2026, OpenAI confirmed its intention to implement an invisible watermarking system for text created by its prominent artificial intelligence models, ChatGPT and Codex, specifically for users within the European Union. This strategic decision comes as a direct consequence of the EU AI Act's transparency regulations, which became effective on August 2, 2026. These regulations necessitate that AI companies clearly mark content produced by their systems, allowing other AI tools to identify its origin.
OpenAI's chosen method, detailed in a technical report titled 'textGrain' co-authored with researchers from the University of Pennsylvania and Yale, involves subtly influencing the AI model's word selection. This process embeds an imperceptible pattern within the generated text, which a specialized detector can identify using a secret key. This innovative approach ensures that the watermark remains embedded even if the text is copied and pasted, without compromising the model's performance or revealing user identities.
The watermarking system will be rolled out to eligible ChatGPT and Codex users in the EU over the coming weeks. Developers worldwide, utilizing OpenAI's API, will also have the option to enable this feature for selected models, although it will not be activated by default. The company has clarified that, at launch, text watermarking will not be a global standard.
However, OpenAI acknowledges certain limitations. Their internal tests indicate that substantial editing, such as replacing just 10% of words with synonyms, can significantly reduce the detection rate from approximately 92% to 66%. Furthermore, shorter passages, mathematical solutions, and translated texts present increased challenges for accurate detection. Consequently, initial access to the detector will be restricted to approved researchers and expert organizations to facilitate thorough evaluation of its reliability and responsible applications.
The company also emphasizes that the absence of a watermark does not definitively confirm human authorship, as text could be too brief, heavily modified, or originate from a different AI system. This initiative follows a similar move by Anthropic, which began watermarking text from its Claude AI model globally two months prior, a decision that generated mixed reactions among its user base. OpenAI had previously developed a text watermark but withheld its release due to concerns about users migrating to rival platforms that lacked such features. This new step underscores a collective industry effort, with major players like Anthropic, Google, Meta, Microsoft, and OpenAI committing to adhere to the EU's code of practice for AI-generated content transparency.
This initiative highlights the growing importance of ethical considerations and regulatory compliance in the rapidly evolving field of artificial intelligence. As AI systems become more integrated into daily life, establishing clear methods for identifying AI-generated content is crucial for maintaining trust and combating potential misuse. While challenges remain in perfecting detection mechanisms, this move by OpenAI represents a proactive step towards responsible AI development and deployment.
