OpenAI details text watermarking strategy in response to EU AI regulations

OpenAI has outlined its approach to text watermarking in compliance with the EU AI Act, which mandates that generative AI providers make generated text identifiable through machine-readable signals. The company plans to introduce text watermarking for select models globally starting today, with the feature opt-in for API customers. In the European Union, eligible ChatGPT and Codex users will receive invisible watermarks over the coming weeks. OpenAI is also opening access to its text watermark detector for approved researchers and expert organizations to evaluate and improve the technology. The text watermarking technology, called textGrain, adds an invisible statistical signal to model outputs. However, the system has limitations, including reduced detection rates for shorter texts and edited content. For instance, at a 1% false positive rate, detection rates for 200-token passages were about 80%, compared to 95% for 40,000-token passages. Editing can also weaken the watermark, with 10% word replacement reducing detection from 92% to 66%. OpenAI emphasizes that watermarks do not establish ownership, verify accuracy, or identify users. The company plans to expand detector access as results become more interpretable and continues to refine its approach in response to evolving standards and regulations.