OpenAI has rolled out textGrain, an invisible watermark for ChatGPT and Codex output, to satisfy the EU AI Act’s transparency rules that took effect August 2. The system nudges word choices against a secret key so a statistical pattern is embedded in generated text. It is rolling out over the coming weeks to eligible ChatGPT and Codex users on all plans, EU only; API developers anywhere can switch it on from October 5, but it stays off by default outside the EU.
Two catches from OpenAI’s own disclosure: detection accuracy is roughly 92%, but falls to 66% when a tenth of the words are swapped for synonyms — a thesaurus defeats it — and it struggles with short text, math answers, and translations. And the detector is only available to approved researchers and expert organizations, so the watermarked public cannot check what it says about text in the wild.
Sources:
- TechCrunch and 9to5Mac reporting (via AI News digest, Oct 6, 2026): https://ainews.malone.gr/
Discussion prompts
- Watermarking without a public detector: is this meaningful transparency under the AI Act, or just a compliance artifact OpenAI can point at when asked?
- Robustness as a design question: if 10% synonym substitution drops detection from 92% to 66%, is any text watermark scheme worth building product detection on top of — or is the arms race already lost?
- Why EU-only by default? Watermarking is opt-in everywhere else. Does geography-limited provenance fragment the ecosystem, or is it pragmatic rollout of a fragile tool?
- Provenance alternatives: C2PA-style cryptographic metadata vs. statistical watermarks vs. nothing — what would you actually trust when verifying who (or what) wrote something?