OpenAI's textGrain watermarks AI text to meet EU rules
ChatGPT and Codex output generated in the EU will carry an invisible word-choice watermark within weeks, and API customers everywhere can now opt in to the same thing.
OpenAI announced textGrain on 5 October, a watermarking system built to satisfy the EU AI Act’s requirements for marking AI-generated text. It works by subtly shaping which words and word-pieces the model picks as it writes, leaving a statistical pattern that a detector holding the right method can pick up later, even after some editing.
Over the coming weeks, ChatGPT and Codex text generated inside the EU will carry this watermark automatically. Outside that, it stays optional: API customers worldwide can opt in for select models, and OpenAI says it will stay off by default and will not become a global default at launch. Detector access is limited for now to approved researchers and expert organisations, not the general public.
Why it matters: if you build on the OpenAI API and need to show content provenance, there is now a real switch to flip rather than a third-party heuristic classifier. It’s also worth knowing the limits going in: OpenAI itself says the watermark can’t identify who used the system or verify accuracy, and heavier edits, paraphrasing or translation can strip it out entirely.
The caveat: a missing watermark proves nothing either way — OpenAI is explicit that absence of a detected signal does not mean a human wrote the text.