OpenAI to watermark AI-generated text in EU

OpenAI plans to introduce text watermarking for AI-generated content in the EU as part of its effort to adhere to the region's stringent AI regulations. The company announced that it will roll out this feature in the coming weeks through its ChatGPT and Codex models. This move is a response to the EU AI Act, which came into effect in August, requiring specific AI systems to disclose when users interact with AI and how the content is generated or altered.

OpenAI's watermarking technology, known as TextGrain, involves adding an invisible statistical signal to the model's word choices. This signal helps the detector identify if a passage contains a watermark. However, the company acknowledged that the technology is still in its early stages and has significant limitations. Detection rates can drop dramatically when the AI-generated content is edited or when only a portion of the words are changed.

OpenAI is one of 200 companies that have signed the EU AI Act's Code of Practice on Transparency of AI-Generated Content, a voluntary initiative aimed at helping businesses comply with the strict regulations. The company also already has several verification mechanisms for other types of content, such as image and audio outputs, which are labeled with C2PA and SynthID watermarks.