Skip to content
TECH STARTUP & BUSINESS NEWS

OpenAI to Introduce Invisible Text Watermarking for ChatGPT and Codex in the European Union to Comply with EU 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 ensure full compliance with the newly enacted European Union AI Act. The company shared the details of the upcoming deployment in a blog post on Monday, outlining how the technology functions, its rollout schedule, and the broader implications for the artificial intelligence industry as regulatory oversight intensifies across major global markets.

The implementation is being driven directly by the transparency rules stipulated under the EU AI Act, which officially took effect on August 2. These regulatory mandates require artificial intelligence developers and providers to ensure that AI-generated content is marked or tagged in a manner that allows other automated systems, platforms, and stakeholders to readily identify its machine-generated origins.

According to OpenAI, the new text-watermarking mechanism will roll out progressively over the coming weeks to all eligible ChatGPT and Codex users across every subscription plan, but the default activation will be strictly restricted to users located within the European Union. Meanwhile, software developers utilizing OpenAI’s application programming interface (API) from anywhere in the world will have the option to manually enable the watermarking feature for select models starting immediately, though it will remain switched off by default for global API users. OpenAI confirmed that it has chosen not to make text watermarking a global default setting at the time of its initial launch.

Rather than appearing as a visible tag, badge, or symbol on the screen, the watermark operates beneath the surface. It works by subtly shaping and steering the underlying language model’s word choices during generation, leaving behind a statistically detectable pattern that remains completely invisible to human readers but easily picked up by a specialized detector. Because this pattern is woven directly into the fabric of the words themselves, it travels seamlessly with the text even when it is copied, pasted, and transferred across different platforms. Furthermore, OpenAI noted that the watermarking process does not track or identify individual users, and internal testing showed no meaningful degradation in the performance capabilities of its models when the feature was switched on.

Alongside the announcement, OpenAI published a comprehensive technical report detailing its watermarking methodology, known as textGrain. Co-authored alongside academic researchers hailing from the University of Pennsylvania and Yale, the report outlines an example of how a secret cryptographic key is used to sort next-word predictions during the sentence-completion process. When hundreds of these subtle statistical nudges are combined within a single passage, an authorized detector can accurately spot AI-generated content using only the text itself and the corresponding secret key.

Despite the sophistication of the technology, questions remain regarding how easily the watermarks can be disrupted or erased through routine text modification. OpenAI’s own rigorous testing suggests that human editing can indeed interfere with detection rates. In one specific evaluation conducted by the company, replacing just 10% of the generated words with alternative synonyms caused the detection accuracy to drop sharply from approximately 92% down to 66%. Additionally, OpenAI acknowledged that short textual passages, mathematical answers, and translated text inherently present greater challenges for reliable detection.

OpenAI will start watermarking ChatGPT’s text in the EU

These limitations heavily influenced the company’s cautious distribution 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, who can assist in evaluating the technology’s overall reliability and exploring responsible use cases.

OpenAI also issued a clear caution regarding the inverse scenario: the absence of a watermark does not constitute proof of human authorship. A piece of text might lack the watermark simply because it is too short, has been heavily edited by a human user, or originated from a completely different company’s AI model, which would use a different watermarking scheme or none at all.

The company emphasized that watermarks can indicate whether an OpenAI system generated or processed a portion of a passage, but they cannot measure the extent of human judgment, creative direction, or subsequent editing that went into shaping the final product.

This development follows a similar industry move made two months prior by rival AI firm Anthropic, which announced that it would begin watermarking text generated by its Claude models on a worldwide scale. That decision sparked notable backlash from a segment of Claude users who complained that the watermarks would catch them bypassing work or academic assignments, arguing that they had supplied all the intellectual instructions, context, and decisions while the AI merely served as a utilitarian tool.

Historical context reveals that OpenAI had previously developed a functional text-watermarking tool in the past but deliberately held off on releasing it to the public. Internal concerns, as reported by The Wall Street Journal in 2024, centered around the fear that users would abandon ChatGPT in favor of competing platforms that did not enforce watermarking restrictions.

As regulatory landscapes shift, OpenAI is far from alone in aligning its compliance strategy with European directives. Major artificial intelligence developers, including Anthropic, Google, Meta, Microsoft, and OpenAI, have all formally committed to adhering to the European Union’s voluntary yet influential code of practice concerning transparency in AI-generated content.