Definition
textGrain is OpenAI’s entropy-calibrated statistical watermarking method for language model text. It subtly shapes token/word choices to embed a detectable signal without visible marks, using optimal transport with Gumbel costs and KL regularization.
Key Points
- 2026-10-05: openai rolls out textGrain for chatgpt and codex in EU over coming weeks; API opt-in globally, off by default (2026-10-05-openai-chatgpt-watermarking-eu)
- Complies with eu-ai-act Article 50 transparency rules effective August 2, 2026
- Detection requires only text + secret key; ~92% detection drops to ~66% after 10% synonym edits
- Short passages, math answers, and translated text harder to detect
- Detector access limited to approved researchers initially; OpenAI plans open-source release
- Does not identify users, verify accuracy, or prove human authorship
Warning
Watermark absence does not prove human authorship. Reliability degrades under editing — similar limitations to synthid-text.