OpenAI prepares invisible watermarking for AI-generated text in the European Union

21:20

OpenAI is preparing to introduce a new system that could make some text produced through ChatGPT and Codex technically identifiable without changing the way it appears to readers. The company says the technology, known as textGrain, will be gradually introduced for eligible users in the European Union over the coming weeks.

The initiative comes as European rules place increasing emphasis on transparency around content created or processed with artificial intelligence. Rather than displaying a visible label, OpenAI's approach embeds a statistical signal directly into the generated text, allowing specialized detection systems to look for evidence of its origin.

For ordinary readers, the difference should be imperceptible. There will be no warning, icon or explicit statement attached to the text indicating that an AI model was involved.

How textGrain embeds its signal

Unlike conventional watermarking techniques that rely on hidden characters or invisible spaces, textGrain operates during the model's generation process. The system subtly influences some word choices, creating a pattern that remains natural to a human reader but can potentially be recognized by a compatible detector.

That design is intended to make the signal more resilient to basic copying and pasting. However, the technology is not designed to provide a complete history of a document.

A successful detection would indicate that text was generated or processed using a compatible OpenAI system. It would not reveal the identity of the user, expose the prompt submitted to the model or establish exactly how much of the final text was subsequently edited by a person.

Rewriting remains a major challenge

The reliability of AI-text detection remains closely linked to the amount of material available and the extent to which the original wording has been changed.

According to OpenAI's reported testing, detection reached roughly 92% under certain conditions for passages of around 400 tokens. Performance fell to about 66% after approximately 10% of the words were replaced with synonyms, and dropped to around 17% when the proportion reached 25%.

Translation and substantial human editing can also weaken the detectable pattern. This means that a failure to identify the watermark cannot, on its own, prove that a text was written entirely by a human.

OpenAI also acknowledges the possibility of false positives, an important limitation for institutions that might consider using automated detection when assessing academic, professional or other written material.

Researchers will initially receive access

The detection technology itself is not expected to be immediately available to everyone. OpenAI plans to provide access initially to selected researchers and specialized organizations so that the system can undergo further testing and evaluation.

The company is also considering broader applications beyond the European rollout. Certain OpenAI API customers will be able to activate the feature on compatible models, although it will remain switched off by default.

OpenAI has additionally indicated that it intends to release textGrain as open-source technology. Such a move could allow independent researchers and developers to examine the approach, test its robustness and explore new applications for AI-content traceability.

A new layer in Europe's AI transparency framework

The deployment of textGrain illustrates a broader shift in the way AI-generated material is being handled in the European Union. As generative AI becomes increasingly integrated into education, business, media and public services, policymakers and technology companies are looking for mechanisms that can distinguish machine-generated content from conventional human writing.

OpenAI's system does not offer a definitive answer to that challenge. Its effectiveness can decline sharply when text is short, translated or substantially rewritten, while detection errors remain possible.

Nevertheless, the technology represents an attempt to move AI transparency beyond visible labels and toward technical signals embedded directly into generated content. If adopted at scale, systems such as textGrain could become part of the emerging infrastructure used to assess the provenance of digital text.