OpenAI announces phased text watermarking for ChatGPT and Codex in the EU
OpenAI announces invisible text watermarks for eligible EU output and API opt-in globally. Detection is limited, and a watermark cannot establish authorship, accuracy or human contribution.
“Does this text contain an AI watermark?” may become a common question. It is a narrower question than whether a person wrote the article.
OpenAI announced a phased watermark rollout for eligible ChatGPT and Codex text in the EU on October 5. API customers worldwide can opt in for selected models; the API default remains off. Detector access initially requires research or expert approval.
The signal is in word selection
The textGrain report describes a statistical relationship introduced during generation using keyed randomness. A detector looks for that relationship in the text. This is not a visible badge or an author’s account name hidden in a corner.
As an explanatory analogy, imagine a detectable tendency in combinations of words rather than a special character. It is statistical evidence, not a guarantee for every passage.
OpenAI says detection can miss signals or report false positives. A result cannot quantify human contribution or identify a user. Failure to detect a watermark is not proof of human authorship.
Read the work, not just the label
Using AI to revise one sentence and asking it to invent an entire article are different processes. A detector result cannot reconstruct that history or assess the quality of the reasoning.
An editorial example: a report with identifiable sources, interview material and dates gives readers something to check, even if its language was edited with assistance. A piece full of unsupported numbers does not become credible because a detector returns a negative result.
The useful question is whether watermarking adds a reliable clue about provenance. It should not become a shortcut for assigning quality or blame. Writers trying to explain how their work was produced can preserve drafts, source material and revision history; those records answer questions that a single score cannot.