OpenAI announces text watermarks for ChatGPT and Codex in the EU
Invisible provenance signals aim to help with AI text. Source checking and an accountable editing process remain essential for publishers.

Record sources alongside the draft.
Assess provenance separately from accuracy.
Review the finished text before publishing.
The short answer
A fluent passage does not make its claims verified. For a publishing team, the practical question is how to document the path from the first AI draft to the final article. A technical provenance signal can contribute to that process; responsibility for the finished text stays with the publisher.
Announcement date: 5 Oct 2026
What changed?
On October 5, OpenAI announced textGrain. Eligible ChatGPT and Codex text in the EU is to receive an invisible statistical signal over the coming weeks. API customers worldwide can opt in for selected models; watermarking remains off by default there.
Who is this relevant to?
This matters to people using AI to write, translate or code. For editors, the useful distinction is between how a passage was produced and whether its content deserves to be published.
What this means for your work
Our editorial suggestion: record original sources and event dates beside the draft. Check names, numbers and availability individually. Note the passages you changed and the claims you removed. Base approval on the checked content. This gives a team a clear account of why an article was published even when a technical signal cannot provide a definite answer.
Limits of this report
The text detector initially serves approved researchers and expert organizations. Detection can be wrong, and editing or translation can weaken it. The signal verifies neither truth nor a particular user identity. We have not tested the detector.
Was this article helpful?
Your feedback helps us improve our articles.
Votes are counted per language version. Repeat voting is limited; totals do not necessarily represent distinct people. Privacy

