OpenAI to apply textGrain watermarks to EU ChatGPT output

OpenAI prepares EU rollout for text watermarks

OpenAI says it will begin adding an invisible watermark to text generated by ChatGPT and Codex for eligible users in the European Union over the coming weeks. The measure applies across plans in the EU and is intended to meet transparency obligations under the EU AI Act, whose relevant rules took effect on August 2.

The company’s method, called textGrain, does not add a visible label or symbol. Instead, it subtly influences the model’s word choices to create a pattern that readers cannot see but a detector can identify. Because the pattern is embedded in the wording, OpenAI says it remains with text that is copied and pasted.

For developers, the company is making text watermarking available from today through the OpenAI API for selected models worldwide. It is disabled by default, and OpenAI is not making it a global default for ChatGPT or Codex at launch.

How textGrain detection works

OpenAI published a technical report with the announcement, co-written with researchers from the University of Pennsylvania and Yale. The report describes using a secret key to sort possible next-word predictions. Repeated choices influenced by that key create enough signals for a detector to identify AI-generated or AI-processed text using the passage and the key.

The EU initiative expands on OpenAI's EU text watermarking rollout while retaining an important limitation: a watermark can indicate that an OpenAI system generated or processed part of a passage, but it cannot establish how much human judgment, editing or creativity contributed to the result.

OpenAI says it found no meaningful change in model performance when watermarking was enabled, and says the technique does not identify the user. However, the company also stresses that an absent watermark is not proof that text was written by a human. A passage may be too short, heavily edited or generated by another provider’s system.

Editing can reduce detection reliability

The company’s tests show that the signal can be diminished through revision. In one test, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Short passages, mathematical answers and translated text are also harder to detect.

Those constraints are why OpenAI plans initially to offer detector access only to approved researchers and expert organizations, which it says can help assess reliability and responsible uses. Anthropic announced two months earlier that it would watermark Claude-generated text worldwide, while Anthropic, Google, Meta, Microsoft and OpenAI have committed to follow the EU code of practice on AI-generated content.

For businesses, watermark results should be treated as a limited provenance signal rather than a conclusive authorship test, and teams using the API should decide whether enabling the optional feature fits their content, review and compliance processes.