textGrain text watermarking
OpenAI announced textGrain, a statistical text watermark. API customers can opt in today, and ChatGPT and Codex text in the EU will be watermarked automatically over the coming weeks.
textGrain nudges the model's word choices so that a long passage carries a pattern that OpenAI's detector can find. It responds to the EU AI Act Article 50 rule on machine-readable marking of AI text. API watermarking is off by default and available worldwide on select models. The detector is open by application to approved researchers and expert organizations only.
- Date
- Monday 5 October 2026
- Lab
- OpenAI
- Kind
- feature
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Detection rate, 200-token passage, at 1% false positive rate | about 80% | company |
| Detection rate, 400-token passage, at 1% false positive rate | about 95% | company |
| Detection rate after 10% of words replaced | 66% down from 92% for unedited text | company |
| Detection rate after 25% of words replaced | 17% | company |
The EU AI Act deadline for existing providers is reported as 2026-12-02. Anthropic announced a similar watermark in August 2026.
Sources
- thenextweb.com/news/openai-text-watermarking-chatgpt-codex-eu-ai-act
- openai.com/index/eu-text-provenance/
This record was checked against its sources on 8 October 2026. How we check