On Monday, October 5, OpenAI announced it would begin adding invisible watermarks to texts generated by ChatGPT and Codex in the European Union (EU). The company said the move is aimed at complying with the transparency requirements of the EU AI Act, with the technology rolling out in stages over the coming weeks.
The AI Act's transparency rules took effect on August 2, requiring AI companies to label generated content so other systems can detect it. In practice, the law requires that other software systems be able to recognize AI-written text by machine. It is in response to this requirement that OpenAI is launching its watermarking method called textGrain. The company also published a technical report on the method alongside the announcement, disclosing its technical details.
How textGrain Works
OpenAI named its watermarking method textGrain. The technical report on the method was written in collaboration with researchers from the University of Pennsylvania and Yale University, explaining how the method works with an example.
According to the report, a secret key arranges the model's likely candidates for the next word in a certain order, and the sentence is completed accordingly. At each prediction, a small bias is applied to the model's word choice — the reader never notices it, but the detector sees the combined pattern of hundreds of such biases. A single biased choice is nearly imperceptible, but repeated across hundreds of sentences it forms an overall statistical pattern. By reading the text and comparing it against that secret key, the detector finds the pattern — no additional information beyond the text itself is needed. It is not a visible mark or a hidden code, but an invisible pattern woven into the text's word choices.
Who Gets It and When
The watermark will be enabled over the coming weeks for eligible ChatGPT and Codex users on all plans in the EU — but only within the EU. The rollout will not happen in a single day: it will proceed in stages over the coming weeks and cover all subscription plans. The company said it does not currently intend to make it a default setting outside the EU.
At the same time, developers working through OpenAI's developer API can, starting today, optionally enable the watermark for certain models anywhere in the world; the setting is off by default. Bloomberg Law also confirmed that the option to enable watermarks for API customers opened on October 5. Two separate paths are thus set out: the watermark is rolled out automatically for users in the EU, while developers working through the API enable it at their discretion. In both cases, it is the same textGrain technology at work.
Editing Weakens the Watermark
According to OpenAI's own tests, the watermark can be weakened through editing. In one test, replacing 10% of the text's words with synonyms dropped the detection rate from about 92% to 66%. The company also noted that short excerpts, math answers, and translated texts are harder to detect. The test results show that the more heavily a text is edited, the less likely the detector is to catch it: a short text or one far removed from its original form blurs the watermark signal.
OpenAI's official chart below shows that the detection rate falls as the amount of editing grows: the detection curve drops noticeably both when 10% of words are replaced and when 25% are replaced.
Who Will Get Detector Access
OpenAI said it will initially grant detector access only to vetted researchers and expert organizations. The detector will not be opened to the general public for now — meaning the watermark-checking tool will remain with a select circle of specialists at the initial stage and will not be available to the wider public. The company also warned that the absence of a watermark in a text does not mean it was written by a human:
"The absence of a watermark does not prove that the text's author is human," OpenAI warned.
The announcement came after Anthropic said two months ago it would watermark texts generated by Claude — Anthropic had said it would apply the measure worldwide. OpenAI's decision, meanwhile, remains limited to the EU for now. With that, OpenAI joins the ranks of companies rolling out text watermarking, following Anthropic.
