OpenAI to Implement Default Watermarking for ChatGPT Outputs in the EU

<p>OpenAI will begin automatically watermarking text generated with ChatGPT in the European Union, as announced by the company. The watermarking feature will also be available in other regions, but it will be disabled by default outside of the EU.</p><p>This initiative in Europe is in response to the EU AI Act, which came into effect in August. The Act mandates that content produced by AI models must be marked in a manner detectable by another tool. Currently, there is no completely effective and reliable method for achieving this. Existing standards, such as SynthID and the C2PA project, are relatively easy to bypass for individuals with basic technical skills.</p><p>OpenAI's watermarking method, referred to as textGrain, is proprietary. The company has published a technical paper detailing its functionality. Generally, it operates similarly to other watermarking tools for large language models (LLMs), embedding patterns in word choices that are not discernible to human readers and do not significantly alter the overall quality of the output. However, these patterns can be detected by someone with the appropriate key using a specialized detector. OpenAI plans to provide access to this detector to a limited number of researchers and organizations, with an approval process for additional requests over time.</p>