OpenAI Rolls Out Invisible Text Watermarking for ChatGPT and Codex in the European Union to Comply with AI Act

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 comply with the stringent transparency requirements mandated by the landmark EU AI Act.

The update, detailed in a company blog post, marks a significant shift for the artificial intelligence giant. The transparency rules outlined in the EU AI Act, which formally took effect on August 2, require major AI developers to ensure that machine-generated content is clearly marked in a way that other systems can reliably identify.

According to OpenAI, the invisible text watermarking feature will roll out progressively over the coming weeks to eligible ChatGPT and Codex users across all subscription plans, though the implementation will initially be restricted to users in the European Union. Meanwhile, developers utilizing OpenAI’s application programming interface (API) anywhere in the world will have the option to manually enable the watermarking feature for select models, though it will remain turned off by default. The company confirmed that it is not making text watermarking a global default setting at launch.

Unlike traditional watermarks that appear as a visible stamp, logo, or metadata tag, OpenAI’s implementation does not rely on an actual symbol. Instead, the mechanism works by subtly shaping the model’s word choices during the text generation process, leaving behind a distinct statistical pattern that remains invisible to human readers but can be easily picked up by a specialized detector. Because the watermark is embedded directly within the vocabulary and syntax of the text itself, it travels seamlessly with the passage whenever it is copied and pasted across different platforms. OpenAI noted that the watermarking process does not identify individual users and emphasized that internal testing showed no meaningful change or degradation in its models’ overall performance when the feature was switched on.

Alongside the commercial announcement, OpenAI published a detailed technical report outlining the methodology behind its approach, which has been named textGrain. Co-written in collaboration with researchers from the University of Pennsylvania and Yale, the report describes how the system uses a secret cryptographic key to sort and bias next-word predictions when completing a sentence. By accumulating hundreds of these subtle textual nudges throughout a longer passage, a designated detector can successfully spot AI-generated content using only the text and the secret key.

Despite its technical sophistication, the watermark is not entirely tamper-proof. When asked whether the watermark can be removed through manual editing or paraphrasing, OpenAI’s testing suggests that it can be diminished. In one internal test, replacing just 10 percent of the words in a watermarked passage with common synonyms caused the detection rate to drop sharply from approximately 92 percent down to 66 percent. Furthermore, the company acknowledged that short passages, mathematical solutions, and translated text are inherently more difficult to detect accurately.

These practical limitations heavily influenced the company’s deployment strategy. OpenAI stated that these constraints contributed directly to its decision to provide initial access to the detection tools exclusively to approved researchers and expert organizations. These trusted groups will assist the company in evaluating the reliability, effectiveness, and responsible uses of the technology before any broader rollout is considered.

OpenAI also issued a broader caution regarding the interpretation of the detection results, noting that a missing watermark does not definitively prove human authorship. A passage might lack a detectable watermark simply because it is too short, heavily edited, or because it originated from a competing company’s AI model rather than OpenAI’s infrastructure. Watermarks can indicate that an OpenAI system generated or processed part of a passage, but they cannot measure how much human judgment, editorial revision, or creative input went into the final product.

This development arrives two months after competitor Anthropic announced that it would begin watermarking text generated by its Claude AI models, a compliance and transparency measure that the company chose to apply worldwide. That decision quickly drew a degree of backlash from some Claude users, who argued that they had supplied the original instructions, conceptual context, and critical decisions while Claude merely acted as a software tool to execute the task.

Historically, OpenAI had developed a text-watermarking system in the past but deliberately held off on releasing it to the public. Reports from outlets like The Wall Street Journal in 2024 indicated that the company harbored concerns that implementing a mandatory watermark would drive users away toward rival platforms that chose not to adopt similar restrictions.

The shifting regulatory landscape in Europe, however, has increasingly forced the industry’s hand. OpenAI, alongside other leading artificial intelligence laboratories including Anthropic, Google, Meta, and Microsoft, is among the major technology companies that have formally committed to following the European Union’s voluntary code of practice on transparency for AI-generated content. As these regulatory frameworks mature, companies are finding it increasingly necessary to balance commercial competitiveness with compliance mandates designed to foster trust and accountability in digital spaces.