OpenAI now offers opt-in textGrain watermarking for selected API models and is phasing it into EU products; detector access remains restricted and editing materially lowers detection.
OpenAI launches opt-in textGrain watermarking for API output
2026-10-05Verified announcementWATCH
OpenAI launched textGrain, an invisible statistical watermark for generated text. API customers can opt in globally for selected models, while eligible ChatGPT and Codex output in the EU will gain watermarks over the following weeks. Detector access is initially restricted to approved researchers and expert organizations. OpenAI plans to publish the implementation, but no public code was available at launch.
Official · ReferenceOpenAI introduces textGrain watermarkingOfficial · PapertextGrain technical paper
Availability: API opt-in for selected models from 2026-10-05; phased EU rollout for eligible ChatGPT and Codex output
What this means for you
For teams with EU provenance obligations, textGrain creates a deployable provider-side signal but not a reliable authorship test. Keep it on WATCH until detector access broadens and independent tests measure false positives, translation and editing resistance on real content.
Evidence & related material
Technical details & scores
| Availability | API opt-in for selected models from 2026-10-05; phased EU rollout for eligible ChatGPT and Codex output |
|---|---|
| Detector Access | approved researchers and expert organizations only at launch |
| Method | entropy-calibrated keyed statistical watermark in token choices |
| Reported Detection | about 80% at 200 tokens and 95% at 400 tokens on psychology passages, at a 1% target false-positive rate |
| Editing Limit | 10% synonym replacement lowered reported detection from about 92% to 66%; 25% lowered it to 17% |
| Public Artifact | technical report and API documentation; implementation promised but not released |
Technical impact 3
Engineering value 3
Adoption signal 3
Maturity 2
Verification cost 3
Risk 3