OpenAI has announced that it will begin invisibly marking texts generated by ChatGPT in the European Union. The measure responds to the requirements of the EU's AI Act, which mandates that content produced by artificial intelligence must be identifiable through automated systems.
The announcement comes at a time when European regulations are beginning to demand concrete traceability mechanisms for AI-generated content. The company led by Sam Altman communicated this on October 5, detailing how the new technology works and, unusually, openly acknowledging its limitations.
How the invisible watermark works
The technology is called textGrain. Unlike other marking systems, it does not insert words, symbols, or additional characters into the text. Instead, it slightly modifies the model's word choice following a statistical pattern that is imperceptible to the human eye.
That pattern functions as a signature based on the rhythm of the writing. It can only be recognized by a specialized detector, which is not available to the general public.
Where and when it will be applied
The rollout of textGrain will not be uniform or immediate. Here are the main details according to OpenAI:
- ChatGPT and Codex in the European Union: the watermark will be incorporated over the coming weeks into texts that meet certain technical requirements. For now, it is not active for all users.
- API for developers, globally: as of October 5, it can be activated voluntarily in some models, although it is disabled by default.
- Restricted access to the detector: initially, only researchers and organizations specifically approved by OpenAI will be able to use it. The company has opened an application form for those who wish to gain access.
- Images and audio unchanged: the existing public tools for verifying whether a visual or audio file comes from OpenAI will continue to work as before.
The limitations that OpenAI itself admits
One of the most notable aspects of the announcement is the frankness with which OpenAI exposes the weaknesses of its system. With a 1% false positive error rate, the detector managed to identify the watermark in approximately 80% of short fragments, equivalent to about 200 units of text, and in 95% of longer fragments, of around 400 units.
The effectiveness drops significantly in technical texts with little room for lexical variation, such as mathematical content, where the model has less freedom to choose between synonyms.
The signal is lost if the text is rewritten
The most revealing data has to do with subsequent manipulation of the content. In a 400-unit text, replacing 10% of the words with synonyms reduced the detection rate from approximately 92% to 66%. If the proportion of changes rises to 25%, detection drops to just 17%.
In practice, this means that a relatively simple rewrite can almost completely eliminate the ability to trace the origin of the text.
OpenAI also maintains that the watermark does not affect the quality of the responses. According to its internal tests with its most recent model, the results with and without the invisible watermark are practically identical. The company has also expressed its intention to release this technology as open source in the future.
What this means for users
The arrival of this technology raises several practical implications that should be kept in mind:
- It is not definitive proof. OpenAI clarifies that the watermark only indicates that the text likely comes from its models, but it does not identify who generated it or which account was used. Nor does a text without a watermark prove that it was written by a person.
- Effect on daily work. Those who use ChatGPT in the European Union to draft content will carry the invisible signal even if they copy the text to another platform. However, a substantial rewrite can remove that mark.
- Caution for teachers and companies. OpenAI itself warns that it is not advisable to use these types of detectors as sole evidence, given that their limitations are significant, especially with edited texts.
Frequently Asked Questions
What is textGrain?
It is the technology that OpenAI uses to invisibly mark texts generated by ChatGPT, subtly modifying the word choice pattern without adding visible characters.
Can this watermark be seen in the text?
No. The text reads exactly the same as any other. The watermark can only be detected with specialized software that, for now, is not publicly accessible.
Why is OpenAI implementing this now?
To comply with the European Union's AI Act, which requires that content generated by artificial intelligence be identifiable through automated tools.
Does it work if I rewrite the text generated by ChatGPT?
The effectiveness decreases considerably. According to OpenAI's data, changing 25% of the words for synonyms reduces detection to just 17%.
Who can use the invisible watermark detector?
For the moment, only researchers and organizations specifically approved by OpenAI, through an application form enabled by the company.
Does this measure affect images or audio generated by OpenAI?
No. The existing public tools to verify the origin of images and audio from OpenAI continue to function without changes.
