Overview
- OpenAI said Monday it has started deploying textGrain, an invisible text watermark that API customers can opt in to worldwide while ChatGPT and Codex outputs for EU users will be watermarked automatically in the coming weeks.
- TextGrain is part of a layered provenance strategy that pairs the invisible text signal with C2PA metadata for files and a licensed SynthID watermark for images and audio.
- Detection reliability depends on passage length and edits, with reported detection rates around 80% for 200-token passages and 95% for 400-token passages at a 1% false-positive rate, but rates fall sharply when words are replaced or text is heavily edited.
- OpenAI is gating access to the watermark detector and opened applications for approved researchers and expert organizations on a case-by-case basis rather than making the tool public.
- The rollout shifts responsibility to developers and deployers to decide when to mark outputs, raises fairness questions for non-native speakers, and leaves regulators to judge whether opt-in API defaults satisfy Article 50 before the December 2, 2026 deadline.