OpenAI to Watermark ChatGPT Text in EU for AI Act Compliance

Overview

OpenAI has announced that it will begin embedding invisible watermarks into text generated by ChatGPT and Codex within the European Union. This strategic move is designed to ensure compliance with the EU AI Act, which mandates that providers of general-purpose AI models take measures to detect and identify AI-generated content. The company detailed the rollout in a blog post released on Monday, marking a significant step toward meeting the bloc’s stringent new regulatory framework.

The watermarking initiative represents OpenAI’s most concrete response yet to the transparency requirements that took effect on August 2. These rules require AI systems to mark their outputs in a manner that other technological systems can recognize. By implementing this feature, OpenAI aims to align its operations with European law while providing a tool that distinguishes machine-generated text from human-authored content for eligible users in the region.

Technical Implementation and Global Rollout

OpenAI stated that the watermark will be deployed over the coming weeks to eligible ChatGPT and Codex users on all subscription plans within the EU. However, the company clarified that this specific feature will not be enabled globally at launch. Instead, it is currently limited to users situated within the European Union. For developers utilizing OpenAI’s API worldwide, the feature is available immediately but is disabled by default. These developers must actively choose to enable the watermark for select models, giving them control over its integration into their own applications.

The technology behind this watermark is distinct from traditional visible tags or metadata. OpenAI described the method, which it has named textGrain, as a technique that subtly shapes the model’s word choices. This process creates a statistical pattern within the text that is imperceptible to human readers but can be detected by specialized software. Because the watermark is embedded in the linguistic structure itself rather than attached as an external file, it persists even when the text is copied and pasted across different platforms or documents.

In conjunction with the announcement, OpenAI published a technical report co-written with researchers from the University of Pennsylvania and Yale University. The report outlines how the system uses a secret key to sort next-word predictions, effectively nudging the model toward specific vocabulary choices. When hundreds of these minor adjustments are aggregated, they form a coherent signal that allows detectors to identify AI-generated content using only the text and the corresponding key.

“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses.”

Limitations and Detection Accuracy

Despite the sophisticated nature of the textGrain system, OpenAI acknowledged several significant limitations in its detection capabilities. The company’s internal tests revealed that the watermark is not immutable. In one specific experiment, replacing just 10% of the words with synonyms caused the detection accuracy to drop precipitously from approximately 92% to 66%. This suggests that simple editing techniques can effectively disrupt the watermark’s signature.

Other factors also complicate reliable detection. OpenAI noted that short passages are inherently harder to analyze because there is insufficient data for the statistical patterns to emerge clearly. Additionally, mathematical answers, heavily edited content, and text that has been translated into other languages present significant challenges for current detection algorithms. The company emphasized that these technical constraints are a primary reason why they are restricting access to their detector tools to vetted researchers and expert organizations for the time being.

Furthermore, OpenAI issued a critical clarification regarding the interpretation of a missing watermark. The company stressed that the absence of a detectable signal does not prove that the text was written by a human. Content might fail to register as watermarked simply because it is too brief, has been substantially modified by a user, or was generated by a different AI provider that does not employ similar watermarking techniques.

“[Watermarks] can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.”

Industry Context and Regulatory Landscape

This announcement follows a similar move by Anthropic, which revealed in July that it would begin watermarking text generated by its Claude model on a global scale. That decision sparked considerable debate among users, some of whom argued that they provided the core intellectual input—such as instructions, context, and decisions—and that Claude merely acted as the mechanical tool. Critics viewed the watermark as an infringement on their agency and ownership of the creative process.

OpenAI had reportedly considered implementing text watermarking previously but hesitated due to competitive concerns. According to reports from The Wall Street Journal in 2024, OpenAI feared that introducing the feature globally might drive users toward rival AI services that did not watermark their outputs, potentially causing a loss of market share. The mandatory nature of the EU AI Act appears to have resolved this dilemma by making compliance a legal requirement rather than a voluntary choice.

OpenAI is not alone in its efforts to align with European regulations. Anthropic, Google, Meta, and Microsoft have all committed to following the EU’s code of practice on AI-generated content. This collective commitment reflects the broader industry’s recognition that voluntary guidelines may no longer suffice under the new legal regime. As the EU’s AI Act continues to shape the global standards for artificial intelligence, companies are increasingly balancing transparency mandates with user privacy and product usability.