Texts generated by ChatGPT will soon have a feature that readers will not be able to detect with the naked eye. OpenAI has announced that it will incorporate an invisible watermark into content generated by its models in the European Union, a measure that will begin to be rolled out in the coming weeks and aims to facilitate the automatic identification of texts created with artificial intelligence.
The technology is called textGrain and works in a different way than traditional watermarks. It does not add symbols, hidden spaces, special characters, or any element that can be appreciated when copying or reading a text. Instead, it statistically modifies the way the model selects certain words, creating a kind of mathematical fingerprint integrated into the writing itself that can later be analyzed by a detector.
A measure linked to European regulations
OpenAI's move comes in response to the transparency requirements of the European Union's Artificial Intelligence Regulation, which includes mechanisms for AI-generated content to be identified by computer systems. The company has chosen to begin this rollout in Europe and, for now, will not make the marking a default setting for all users worldwide.
The system will arrive at ChatGPT and Codex in the European Union, with a progressive rollout over the next few weeks and regardless of the contracted plan. At the same time, OpenAI API customers worldwide will be able to voluntarily activate the marking in certain models, although it will remain deactivated by default in the API.
The mark will not be infallible
OpenAI also recognizes that textGrain has limitations. Detection works better the longer the text and can lose accuracy when a person substantially modifies the wording. In its tests, the company found that replacing about 10% of the words in a 400-token text with synonyms reduced detection from approximately 92% to 66%. If the changes reached 25%, the rate dropped to 17%.
There is also the risk of false positives and negatives. In OpenAI's tests, with a target false positive rate of 1%, the detector identified the watermark in approximately 80% of 200-token texts and reached 95% in 400-token texts in certain content. The company therefore emphasizes that the tool should not be interpreted as an absolute proof of authorship.
And there is another important issue: the watermark will not allow knowing who used ChatGPT, what instructions the model received, or what conversation was behind the text. Nor will it determine how much work a person contributed or serve to establish who owns the content. Even if a detector does not find the watermark, it would not necessarily prove that a text was written entirely by a person.
OpenAI plans to start by giving access to its detector only to researchers and specialized organizations, with the aim of studying its operation and improving the technology before expanding its availability. The company also assures that it is working to convert textGrain into open-source technology.