OpenAI is introducing machine-readable watermarking for text generated by ChatGPT and Codex in the European Union. This initiative aligns with the EU AI Act, which mandates that generative AI providers ensure content is identifiable. The rollout will occur over the coming weeks across all subscription plans.
The technology, named textGrain, embeds a statistical signal into text by influencing word selection. A specialized detector then identifies this signal. OpenAI notes that detection accuracy depends on text length and content type. In tests, the system detected watermarks in 80% of 200-token texts and 95% of 400-token texts at a 1% false-positive rate.
The company warns that the technology has limitations, particularly regarding edited content. Replacing 25% of words in a 400-token text reduced detection rates to 17%. Consequently, OpenAI is not releasing the detector publicly yet but is accepting applications from researchers and expert organizations to evaluate its reliability.
Globally, OpenAI is launching optional text watermarking for API users starting today. While this feature is disabled by default, developers can enable it for specific models. The company also plans to open-source the technology and extend support to models available through cloud partners in the near future.
