OpenAI rolls out invisible text watermarking for ChatGPT and Codex in the EU

OpenAI announced that it will embed an invisible, machine-readable watermark named textGrain into the text generated by ChatGPT and Codex, starting with eligible users in the European Union. The rollout is limited to the EU to comply with the region’s AI Act, and the company does not intend to make it a global default at launch, allowing time to gather real-world feedback. API customers worldwide may now choose to enable watermarking for selected models, and OpenAI is working with cloud partners to extend the feature to their services.

Researchers and expert organisations can apply for access to a detection tool that reports the presence of an OpenAI watermark without revealing user data. While internal benchmarks indicate that watermarked text performs on par with standard output, OpenAI cautions that the watermark does not ensure reliable detection, ownership verification, or proof of human contribution.