OpenAI announces textGrain watermark for EU ChatGPT and Codex

OpenAI’s internal evaluation shows the detector works best on longer English texts. For a 400‑token passage, replacing 10% of the words with synonyms lowered detection from roughly 92% to 66%, and a 25% replacement cut it to 17%. Under a 1% false‑positive target, detection was about 80% for 200‑token psychology texts and 95% for 400‑token texts, but rates fell for domains with limited word choice, such as mathematics. The company stresses that a detected watermark only indicates that its system was involved; it does not reveal the amount of human editing, ownership, or accuracy, nor does the absence of a watermark prove human authorship.

OpenAI is limiting detector access to approved researchers and specialized organizations while layering text provenance with other signals like Content Credentials and SynthID for images and audio.