OpenAI to watermark ChatGPT and Codex text in the EU under AI Act provenance rules

OpenAI published its approach to EU text provenance rules: invisible watermarks will be applied to ChatGPT and Codex output for EU users on all plans over the coming weeks. API developers anywhere can optionally enable watermarking on select models starting now. Access to detection tools begins with researchers before any broader rollout.

The post details where watermarks apply and how detection works, a direct response to EU AI Act transparency obligations for generative systems.

This makes OpenAI one of the first major labs to ship text watermarking at consumer scale, a technique that has been researched for years but rarely deployed because of robustness and user-backlash concerns.

Skeptics point to limits: reported detection fails under 200 tokens and drops from 92% to 66% accuracy when 10% of words are replaced, so light paraphrasing defeats it. Watch whether rivals follow, whether EU regulators accept this level of robustness, and whether users migrate to unwatermarked tools.