OpenAI will begin adding an invisible watermark to texts created by ChatGPT and Codex for users in the European Union. The company says this is due to the requirements of the EU AI Act: the new rules require developers to create ways that allow other systems to recognize AI-generated content.

The labeling will appear for ChatGPT and Codex users in the EU within the next few weeks, regardless of subscription tier. At the same time, it will not become a global standard for now: developers using the OpenAI API outside the European Union will be able to enable the feature themselves for some models, but it will remain disabled by default.

The watermark is hidden inside the text itself

There will be no visible symbol or special note in the text. OpenAI uses a method called textGrain, which subtly adjusts the model’s word choice.

When generating text, the model chooses the next word each time from many possible options. OpenAI’s method introduces small changes into this process, creating a statistical pattern that humans do not notice, but that a special detector can identify.

If enough of these changes accumulate, the system will be able to determine that the text was created or processed by an OpenAI model. At the same time, the marker remains inside the text itself and therefore survives ordinary copying and pasting.

OpenAI claims that the watermark does not make it possible to identify the user and has virtually no effect on model quality.

The watermark can be partially erased

The main problem with this approach is that the labeling is not invulnerable.

In its own tests, OpenAI checked what would happen if the text were edited and some words replaced with synonyms. Replacing just 10% of the words reduced the probability of detecting AI-generated text from about 92% to 66%.

In addition, the watermark is harder to detect in short passages, mathematical answers, and translations.

Therefore, the absence of a marker does not mean that a human wrote the text. It may have been too short for reliable analysis, heavily edited, or even created by another AI system altogether.

OpenAI also emphasizes another limitation: even a detected watermark does not show exactly what role artificial intelligence played in creating the text. It may indicate that OpenAI generated or processed part of the material, but it does not say how many human decisions, edits, or how much creativity went into the final result.

OpenAI will not make the detector available to everyone yet

Because of these limitations, the company does not plan to immediately open access to the watermark detection tool for all users. At the first stage, only approved researchers and specialized organizations will get access to it.

OpenAI hopes to use this approach to test the system’s reliability and explore options for its responsible use.

The company described the textGrain method in a separate technical report prepared together with researchers from the University of Pennsylvania and Yale University. It describes the use of a secret key with which the system sorts possible next words in a certain way during text generation. A large number of small changes ultimately forms a recognizable statistical trace.

Why OpenAI is doing this only now

OpenAI had already developed text-labeling technology, but previously chose not to release it. One of the reasons was concern that users would simply switch to competitors whose models do not leave such markers.

Now the situation has changed because of European legislation. The transparency requirements of the EU AI Act took effect on August 2 and require content created by artificial intelligence to be labeled in a way that allows other systems to recognize it.

OpenAI is not the only company moving in this direction. Anthropic has already announced global labeling of texts created by Claude. In addition, OpenAI, Anthropic, Google, Meta, and Microsoft have joined the European code of practice on AI content transparency.

At the same time, OpenAI’s approach highlights the main difficulty of text watermarks: it is relatively easy to determine the origin of long and nearly unchanged AI-generated text, but the more heavily a person edits the material, the shorter it is, or the more specialized structure it contains—such as mathematical formulas—the less reliable detection becomes.

Therefore, OpenAI’s new system is more likely to become a tool for confirming the origin of certain texts than a universal way to determine whether a piece was written by a human or by artificial intelligence.