OpenAI to Automatically Watermark ChatGPT Text in EU

OpenAI has announced a new feature where ChatGPT-generated text will be automatically watermarked in the European Union. While this functionality is also available in other regions, it is turned off by default outside the EU.

This decision in Europe is in response to the EU AI law that came into effect in August. The law requires content generated by an AI model to be marked in a way that can be detected by another tool. Although existing standards like SynthID and the C2PA project exist, they are still relatively easy to bypass for individuals with basic knowledge.

OpenAI’s watermarking method, known as textGrain, introduces unique patterns into the word selection of the generated text. While the watermark is designed not to impact the quality of the output significantly, it can be detected by specialized detectors using a key. OpenAI will provide access to the detector to select researchers and organizations, with plans to expand access over time through an approval request process.

Learn more about textGrain here