Overview
- OpenAI disclosed on Monday that it will roll out textGrain, an invisible, machine‑readable watermark, to eligible ChatGPT and Codex outputs in the European Union over the coming weeks.
- API customers worldwide can opt in to textGrain for select models starting now, but the API setting is off by default and OpenAI is not making the watermark a global default at launch.
- textGrain works by nudging word choice with a secret key to leave a statistical pattern a detector can spot, and OpenAI plans to publish a technical report and open‑source the approach over time.
- OpenAI’s tests show detection is strong on long, unedited passages (about 80% for 200‑token and 95% for 400‑token samples at a 1% false‑positive target) but falls sharply after editing, translation, short length, or constrained language use.
- Access to OpenAI’s detector is limited at launch to approved researchers and expert organizations, and the company warns a detected watermark only signals OpenAI model involvement while a missing watermark does not prove human authorship.