What OpenAI Announced

In a blog post published on October 5, 2026, OpenAI said it will start adding an invisible watermark to text generated by ChatGPT and Codex. The decision was taken to comply with the European Union's Artificial Intelligence Act (AI Act). According to the company, the marking will roll out in the coming weeks to ChatGPT and Codex users on all plans in the EU.

Developers worldwide can already enable the marking for selected models, but outside the EU it will remain off by default. In other words, automatic watermarking currently applies only within the EU. Ars Technica independently confirmed the report, noting that default-on watermarking applies only in the EU, consistent with the company's statement. The watermarking covers relevant ChatGPT and Codex users on all plans: the limitation applies on a territorial basis, not by plan tier. The rollout will not be instant but phased — over the coming weeks. It is one of the major steps by AI labs toward making AI-generated text detectable at scale.

How the Watermarking Works

The method is called textGrain. OpenAI published a technical report on it; the report was written together with researchers from the University of Pennsylvania and Yale University and openly describes how the method works. Using a secret key, textGrain orders the model's predictions for the next word: as text is generated, word choices are subtly aligned with a predefined pattern.

As a result, readers cannot perceive any difference in the text, but a dedicated detector can recognize the pattern. The mark travels with copied text: it is preserved even when a fragment is pasted elsewhere. Crucially, the mark does not identify the user — it only indicates that the text was generated by AI. According to OpenAI, watermarking has had no noticeable impact on model quality. The technical report explains that the secret key orders the model's next-word predictions in a specific order, and that order forms a detector-readable pattern: the text stays natural to readers while a hidden mark is embedded in its statistical structure.

"OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act, the company said Monday in a blog post." — TechCrunch report

Detection Limits

Watermarking is not a perfect defense. In its own tests, OpenAI noted that replacing 10% of words with synonyms reduced detector accuracy from about 92% to 66%. Short fragments, math answers, and translated text are also harder to detect.

For that reason, the detector will not initially be open to the general public: access will be granted only to vetted researchers and expert organizations. The company has openly disclosed the limitations and presents the technology as a practical tool on the path to detecting AI text. The watermarking's real-world effectiveness will be judged exactly within those limitations.

Who Gets Detector Access

Access to the detector is initially restricted: only vetted researchers and expert organizations will receive it. The company explains this approach by the need for independent evaluation of the technology. The accuracy figures — including the drop from 92% to 66% with synonym replacement — will be checked during that evaluation.

So the marking system has two parts: applying the mark is automatic and mass-scale (in the EU), while reading the mark currently rests with a narrow circle of specialists. OpenAI has not yet said when it will open the detector to the public.

OpenAI linked the move directly to the EU AI Act. The law requires that content coming from AI systems be detectable — which is why the company is enabling watermarking by default specifically for EU users. Outside the EU, watermarking remains voluntary: developers can already enable it for selected models, but it stays off by default. So inside the EU the mark is applied automatically, while elsewhere the decision stays with the developer.

The timeline: OpenAI announced the plan in a blog post on October 5, and the next day Ars Technica independently confirmed the report. Practical implementation will be phased, completing over the coming weeks. Ars Technica's independent confirmation recorded precisely this aspect — default-on watermarking only in the EU — along with the company's explanation of legal compliance. The decision is a clear example of how major AI labs are deploying content-detection mechanisms under legislative requirements.