Overview
- OpenAI published textGrain on Monday, October 5, 2026, and began a phased rollout that lets API customers opt in to watermarking today while eligible ChatGPT and Codex outputs in the European Union will be watermarked automatically in the coming weeks.
- textGrain works by subtly nudging a model’s word choices to embed a hidden statistical pattern that a detector can identify using a secret key, and the pattern travels with the text when it is copied or pasted.
- OpenAI published measured performance numbers showing the detector finds roughly 80% of watermarked 200‑token passages and about 95% of 400‑token passages at a 1% false‑positive target, while edits, translation, short passages, and code can sharply reduce detectability.
- Access to OpenAI’s watermark detector is initially limited to approved researchers and expert organizations because the company says the tool can produce false positives and misses and needs further evaluation.
- The rollout shifts practical responsibility to developers and services that use OpenAI models, offers a provenance signal rather than proof of authorship or accuracy, and follows similar watermark moves by rivals such as Anthropic and Google DeepMind.