OpenAI is introducing a default watermarking system for ChatGPT outputs specifically for users in the European Union. This decision follows the implementation of the EU AI Act in August, which mandates that AI-generated content must be identifiable by detection tools.
The proprietary technology, named textGrain, embeds patterns into word choices that remain imperceptible to human readers. These patterns can be identified by a specialized detector, to which OpenAI will grant access to a limited group of researchers and organizations, with a formal request process for broader access.
Internal testing indicates a 92 percent detection success rate under optimal conditions. However, the effectiveness drops significantly if the text is altered: modifying 10 percent of the content reduces detection by nearly 30 percent, while a 20 percent change lowers the success rate by approximately 75 percent.
The company notes that detection performance is generally lower for translated or short texts. While competitor Anthropic has implemented watermarking on a global scale, OpenAI is currently limiting the default rollout to the EU for its ChatGPT and Codex applications, while keeping it optional for API users.
