OpenAI is preparing to introduce invisible text watermarking for eligible ChatGPT and Codex users across the European Union. The rollout is designed to support the transparency requirements of the EU AI Act, which call for AI-generated material to be identifiable by other systems.
Rather than adding a visible label, the approach subtly guides a model's word-selection patterns. These small statistical signals create a hidden signature that can travel with text when it is copied or shared, while remaining imperceptible to readers.
How the textGrain system works
OpenAI's method, called textGrain, uses a secret key to organize likely next-word choices during generation. Across a longer passage, these controlled choices form a pattern that specialized detection tools can identify.
The company developed and documented the system with researchers from the University of Pennsylvania and Yale University. OpenAI says enabling the watermark does not produce a meaningful reduction in model performance and does not reveal the identity of an individual user.
The feature will be introduced in the coming weeks for qualifying ChatGPT and Codex accounts in the EU, regardless of plan type. Developers globally can also choose to activate it for selected models through OpenAI's API, where it will initially remain optional.
Detection with clear limits
OpenAI notes that text watermarking is a provenance signal rather than definitive proof of authorship. Detection can become less reliable when writing is heavily revised, translated, shortened, or built around mathematical responses. In company testing, replacing around 10% of words with synonyms reduced detection accuracy from roughly 92% to 66%.
For the initial phase, detector access will be limited to approved researchers and expert organizations, allowing reliability and responsible applications to be assessed. The initiative points toward a future where AI-assisted writing can be understood with greater context, supporting more informed and transparent digital communication.
