OpenAI to introduce invisible watermark in ChatGPT and Codex texts in the EU

OpenAI announced on Monday, October 5, 2026, that it will introduce a statistical watermark in texts generated by ChatGPT and Codex in the European Union. The feature, called textGrain, will be optional for API customers in any country, but will not be enabled by default.

How the watermark works

During text generation, the model selects tokens based on probabilities. textGrain changes this process by using values linked to a secret key and the context of the passage, creating a recognizable pattern that indicates OpenAI’s involvement.

The detector compares the final text with the same key and checks whether the word choices follow the expected pattern more frequently than by chance. The mark does not add hidden characters, invisible spaces, or unusual punctuation.

We are expanding our content provenance approach to include text, in response to EU regulatory requirements, recognizing the significant limitations of current text marking technology.

Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model.

In the EU, we will begin marking eligible ChatGPT and Codex texts in the coming weeks to comply with the European AI Act.

Customers using our API can enable text marking for selected models worldwide starting today.

— OpenAI (@OpenAI) October 5, 2026

Limitations and detection performance

In internal tests, with a target false positive rate of 1%, the detector identified the mark in about 80% of 200-token passages and 95% of 400-token passages in content such as psychology. Results were lower in mathematics, where there is less flexibility in word choice.

Another evaluation showed that replacing 10% of words with synonyms reduced detection from approximately 92% to 66%. With 25% replacement, the rate dropped to 17%. The tests were conducted in English.

These percentages reflect specific conditions and do not guarantee performance on any text, language, or situation.

European regulatory context

The measure is part of OpenAI’s response to the AI Act, the European Union’s regulation that requires providers to make AI-generated content identifiable in machine-readable format, taking into account the technical limitations of each medium.

The European Commission differentiates this marking from the obligations of those who publish or use the material. For texts that inform the public about matters of public interest, the rules provide an exception when there is human review and editorial responsibility.

Global perspectives and other companies

Although the watermark will not be implemented as a global standard, API customers can enable it for selected models in any country, including Brazil. The feature will remain disabled by default.

Anthropic has already announced plans to introduce watermarking for Claude’s texts. OpenAI plans to make textGrain available as open-source code.

The detector will not be publicly available at launch. Researchers and specialized organizations may request access, subject to approval.