OpenAI’s textGrain Watermarks ChatGPT and Codex Text in the EU — API Opt-In Worldwide
OpenAI will roll out textGrain, an invisible statistical text watermark, to eligible ChatGPT and Codex users on all plans in the EU over coming weeks to meet AI Act transparency rules. API customers worldwide can enable it per project or organization now; it stays off by default — unlike Anthropic’s global Claude default. Light paraphrase defeats detection; a missing watermark proves nothing. The desk does not claim generated code is watermarked.
Brussels’ transparency rule is changing how frontier chat products behave in Europe — and OpenAI chose a narrower default than Anthropic.
On October 5, 2026, OpenAI said it will roll out textGrain — an invisible statistical watermark built into the model’s word choices — to eligible ChatGPT and Codex users on all plans in the EU over the coming weeks, to comply with EU AI Act transparency rules that took effect Aug. 2. API customers worldwide can enable it per project or organization for select models starting now; it stays off by default. That is the opposite of Anthropic’s Claude text watermark, which Anthropic applied globally because it said it could not yet scope by region.
How it works — and where it fails
textGrain nudges next-token choices under a secret key so a detector can spot a pattern in longer passages. OpenAI says it does not identify the user, adds no hidden characters, and showed no meaningful quality or speed change in its tests (including on Astra). Detection access is limited to approved researchers (named partners include Cornell’s John Thickstun, ETH Zurich/INSAIT’s Martin Vechev, and KInIT). OpenAI plans to open-source textGrain and published a technical report with Penn and Yale co-authors.
OpenAI’s own figures (via its Help Center charts as read by TechCrunch and The New Stack): about 80% detection of 200-token and 95% of 400-token passages at a 1% false-positive rate in domains like psychology. Replacing 10% of words cut detection from ~92% to 66%; 25% replacement: 17%. Short text, math, code and translations are harder. Across official EU languages at 1% FPR, detection ranged from 69.0% (Spanish) to 42.2% (Romanian) before strength tuning.
OpenAI itself says a missing watermark does not prove human authorship — the text may be too short, heavily edited, or from another model. Light paraphrase is enough to degrade the signal.
Code caveat
OpenAI’s Help Center states code is harder to watermark because there are fewer plausible next tokens, and the EU Code of Practice does not require watermarks for code snippets or outputs under ~200 tokens. The New Stack notes OpenAI has not clarified what “eligible” Codex output means or whether generated source code itself carries the mark. This piece does not claim code is watermarked.
Limits
- Primary materials inspected: OpenAI Help Center provenance FAQ and TechCrunch / The New Stack secondaries. OpenAI’s separate “Our approach to EU text provenance rules” blog was not fetched as a standalone page.
- Detection rates and edit-sensitivity figures are OpenAI’s (charts summarized by TNS/TechCrunch).
- EU ChatGPT/Codex rollout is “coming weeks,” not a dated GA day.
- API watermarking is opt-in and model-selective; detector access is gated separately.