OpenAI Begins Implementing Watermarks for ChatGPT Text in the European Union
The company is deploying an invisible watermarking system for ChatGPT and Codex texts to comply with the transparency rules of the EU AI Act.
OpenAI has announced that it will begin adding an invisible watermark to text generated by ChatGPT and Codex for users in the European Union. This move aims to meet the transparency requirements of the EU AI Act, which entered into force in August.
New Implementation Under the EU AI Act
In a published blog post, OpenAI announced that it is taking a new step to comply with legal regulations in the European Union. The transparency rules of the act require artificial intelligence companies to mark their content in a way that other systems can recognize.
Use Cases and Status for Global Developers
The new watermarking feature will be rolled out to eligible ChatGPT and Codex users across all plans in the European Union in the coming weeks. Developers worldwide can enable this feature for specific models via the OpenAI API starting today, though the feature will be disabled by default.
Technical Infrastructure of the Invisible Watermark
The watermark does not appear as a visual symbol; instead, it subtly shapes the model's word choices to leave a hidden pattern that readers cannot see but detectors can spot. This way, even if the content is copied and pasted, the watermark travels along with the text.
The textGrain Method and Test Results
OpenAI also shared a technical report titled "textGrain," co-authored by researchers from the University of Pennsylvania and Yale. In tests conducted, it was observed that replacing 10% of the words with their synonyms reduced the detection rate from 92% to 66%.