textGrain adds an invisible statistical signal through the model's word choices, and OpenAI's detector looks for that signal to assess whether text contains an OpenAI watermark. OpenAI explicitly says the watermark does not identify a user, account, prompt, or conversation and does not establish accuracy, ownership, or the amount of human contribution. Detection also becomes less reliable after rewriting, translation, or word substitution.
For the API, global opt-in begins on the announcement date, remains off by default, and is limited to select models. For ChatGPT and Codex, the rollout is EU-only at launch: eligible text output across all plans will gain watermarking over the coming weeks. OpenAI says it is not making text watermarking a global default at launch. Detector access is also not public initially and is limited to approved researchers and expert organizations.