OpenAI has announced that it will begin embedding invisible watermarks into text generated by ChatGPT and Codex within the European Union, a move designed to ensure full compliance with the newly enacted EU AI Act. The decision marks a significant shift for the artificial intelligence pioneer, which had previously hesitated to roll out text-provenance tools globally due to competitive pressures and commercial concerns.
According to a company blog post published on Monday, the transparency rules mandated by the EU AI Act officially took effect on August 2. These regulatory standards require major AI developers and providers to mark AI-generated content using methodologies that enable other systems, platforms, and automated tools to accurately identify it.
The invisible watermarking system will roll out gradually over the coming weeks to eligible users of ChatGPT and Codex across all subscription tiers, though the deployment will be restricted exclusively to the European Union for general users. Meanwhile, developers leveraging OpenAI’s API can choose to enable the watermarking feature for select models globally starting immediately, though it will remain switched off by default outside of regulatory mandates. OpenAI confirmed that it is not making text watermarking a global default setting at the time of its initial launch.
Rather than relying on a visible symbol, badge, or external metadata tag, the watermarking mechanism works by subtly shaping the model’s word choices during generation. This leaves behind a statistical pattern that remains invisible to human readers but can be readily picked up by specialized detectors. Because the watermark is embedded directly within the vocabulary and syntax of the text itself, it travels seamlessly with the passage even when it is copied, pasted, transferred, or shared across different platforms. Furthermore, OpenAI stated that the watermarking process does not track or identify individual users, and internal testing revealed no meaningful degradation in model performance or output quality when the feature is enabled.
Alongside the announcement, OpenAI released a comprehensive technical report detailing its watermarking methodology, which is called textGrain. Co-written in collaboration with academic researchers from the University of Pennsylvania and Yale University, the report walks through the underlying mechanics of the system. The approach uses a secure, secret key to influence and sort next-word predictions as the model completes a sentence or paragraph. When hundreds of these subtle statistical nudges are accumulated across a longer passage, a designated detector can accurately spot AI-generated content using only the text in question and the corresponding secret key.
Despite the sophistication of the textGrain approach, OpenAI acknowledged several inherent limitations regarding the robustness of the watermarking system, particularly when text undergoes human or automated editing. To evaluate how edits impact detection rates, the company conducted various tests, which suggested that watermarks can indeed be disrupted or removed. In one specific test, replacing just 10 percent of the generated words with standard synonyms caused the detection rate to drop significantly, falling from approximately 92 percent down to 66 percent. Additionally, the company noted that shorter passages, technical math answers, and translated text inherently present greater challenges for reliable detection.
These technical boundaries directly informed the company’s deployment strategy. OpenAI explained that these limitations contributed to their decision to provide initial detector access exclusively to approved researchers, academic institutions, and expert organizations who can assist in evaluating the technology’s reliability, resilience, and responsible use cases.
OpenAI also issued broader cautions regarding the interpretation of the watermarking tool, emphasizing that a missing watermark does not serve as definitive proof of human authorship. A passage might lack a detectable watermark simply because it is too brief, heavily modified, or because it originated from an entirely different artificial intelligence system developed by a competing organization.
Watermarks can ultimately indicate that an OpenAI system generated or processed at least a portion of a given passage, but they cannot measure how much human judgment, editorial oversight, editing, or creative input went into the final product, according to the company.
This policy shift by OpenAI follows a similar move made two months prior by rival AI lab Anthropic, which announced that it would begin watermarking text generated by its Claude models on a worldwide basis. Anthropic’s global rollout drew immediate and vocal backlash from some Claude users, who argued that they had supplied all the core instructions, creative context, and decision-making processes while the AI merely acted as a utility tool.
The debate over text provenance touches on long-standing industry tensions. Reports from outlets like The Wall Street Journal indicated in 2024 that OpenAI had previously developed a functional text-watermarking tool but deliberately withheld its release. Internal discussions at the time revealed concerns that implementing the feature could disadvantage OpenAI if users migrated to rival platforms that did not enforce similar watermarking restrictions.
With the implementation of the EU AI Act, the competitive landscape has shifted toward regulatory compliance. OpenAI, alongside other leading industry players such as Anthropic, Google, Meta, and Microsoft, has formally committed to following the European Union’s broader code of practice concerning the transparent labeling and identification of AI-generated content.