OpenAI Will Automatically Watermark ChatGPT Text in the European Union

OpenAI has announced that ChatGPT-generated text will be automatically watermarked in the European Union. Watermarking is also available in some other regions, but it remains turned off by default outside the EU.

EU AI law drives ChatGPT watermarking

The change is intended to comply with the EU AI law, which came into force in August. Under the law, content generated by an AI model must be marked in a way that another tool can detect.

However, there is still no completely effective or reliable method for identifying AI-generated text. Existing standards, including SynthID and the C2PA project, can be relatively easy to bypass for people with basic technical knowledge.

How OpenAI’s textGrain watermark works

OpenAI’s watermark appears to face similar challenges. The company has developed a method called textGrain and published technical documentation explaining how it works.

Like other large language model watermarking systems, textGrain introduces patterns into the model’s word choices. These patterns are not obvious to human readers and are designed not to significantly affect the quality of the text. A specialized detector with the correct key can then identify the watermark.

Who will be able to detect the watermark?

OpenAI says it plans to give access to its textGrain detector to a limited number of researchers and organizations. It will also create an approval process for additional researchers and organizations seeking access over time.