OpenAI Launches textGrain Watermarking to Meet EU AI Act

OpenAI is introducing textGrain, an invisible, machine-readable watermarking system for text generated by ChatGPT and Codex. Developed in collaboration with researchers from Yale University and the University of Pennsylvania, the technology subtly influences word choices to create statistical patterns detectable via a secret key. The rollout begins over the coming weeks for eligible users across all plans within the European Union to comply with the EU AI Act. Globally, API users can opt in to activate the feature for select models starting October 5, though it remains off by default for consumer apps.

Technical tests reveal significant limitations in the system's efficacy. Detection rates dropped from 92% to 17% when 25% of the words were replaced with synonyms. The system also struggles with short passages, mathematical content, and translated text. Because of these risks and the potential for false positives, OpenAI is limiting the detection tool to approved researchers and expert organizations rather than the general public.

This implementation follows a similar worldwide rollout by Anthropic for its Claude models in August. OpenAI noted that while the watermark indicates a system processed the text, it does not prove human authorship, establish ownership, or verify the accuracy of the content.