OpenAI announced it will start embedding an invisible watermark in text generated by ChatGPT and Codex for users in the European Union to comply with the EU AI Act’s transparency rules, which took effect on August 2. These rules require AI-generated content to be marked in a way that other systems can identify.
The watermark will be rolled out over the coming weeks to eligible ChatGPT and Codex users across all subscription plans within the EU. Developers worldwide using OpenAI’s API can enable the watermark for select models starting immediately, though it remains off by default outside the EU. OpenAI has no plans to make watermarking a global default at this time.
Unlike a visible symbol, the watermark works by subtly influencing the model’s word choices, creating a pattern detectable by specialized tools but invisible to readers. Because the watermark is embedded in the text itself, it remains intact when the text is copied or pasted. OpenAI emphasized that the watermark does not reveal user identity and does not significantly impact model performance.
Alongside the announcement, OpenAI released a technical report detailing its watermarking method, called textGrain. Developed in collaboration with researchers from the University of Pennsylvania and Yale, the method uses a secret key to reorder next-word predictions, embedding hundreds of small signals that collectively enable detection of AI-generated content.
OpenAI acknowledged that the watermark can be weakened or removed by editing. Tests showed that replacing 10% of words with synonyms reduced detection accuracy from about 92% to 66%. Detection is also more challenging for short passages, mathematical answers, and translated text. Due to these limitations, OpenAI is initially granting detector access only to approved researchers and expert organizations to assess reliability and responsible use.
The company also cautioned that absence of a watermark does not confirm human authorship, as heavily edited or short texts may evade detection, and other AI systems may produce unwatermarked content. OpenAI stated that watermarks indicate AI involvement but do not measure the extent of human input or creativity.
This move follows Anthropic’s announcement to watermark text generated by its AI model Claude globally, a decision that faced criticism from some users who felt their contributions were underrepresented. OpenAI had previously developed watermarking technology but delayed its release partly due to concerns about losing users to competitors without watermarking.
OpenAI joins other major AI companies, including Anthropic, Google, Meta, and Microsoft, in committing to the EU’s code of practice on AI-generated content transparency.