October 12, 2026, 5:57 am | Read time: 3 minutes
OpenAI will gradually embed invisible watermarks in texts from ChatGPT and the coding assistant Codex in the European Union. The technology, called textGrain, creates a hidden pattern by selecting certain words, which a verification program can detect. For readers, the marking remains invisible. However, the control is not reliable, especially if texts have been subsequently altered or contain only a few words.
Hidden Pattern in Word Choice
Whether it’s an application, complaint, or longer article–texts created by ChatGPT are to contain a machine-detectable trace in the future. OpenAI plans to gradually introduce the watermarks within the European Union in the coming weeks.
The technology is to be used not only with ChatGPT but also with the coding assistant Codex. textGrain influences word selection to create a hidden pattern in the text. This structure is not recognizable during normal reading.
A special verification program can search for the pattern and thus provide clues about the origin of a text. The background is the European AI Act, which requires providers like OpenAI to label AI-generated texts so that computers can recognize the marking.
Better AI Images? OpenAI Expands Features With Update
ChatGPT to Evolve into a Super App
Verification Program Initially Limited
Users currently cannot verify themselves whether a text contains such a watermark. OpenAI initially provides the verification program only to approved researchers and professional organizations. Interested professionals have been able to apply for access since October 5.
Developers can also gain access to the detection depending on the model used. This applies, for example, when ChatGPT is integrated into their own programs via an application programming interface, or API.
In connection with handling personal information, the free tool “AI Paper Trail” is also highlighted. It aims to reveal what personal data conversations with an AI can disclose about users.
Also of interest: ChatGPT Becomes a Fashion Advisor
Detection Loses Security Through Changes
OpenAI justifies the restricted access with potential errors in verification. The program can overlook existing watermarks. Conversely, detection is also possible even though an examined text does not actually contain the corresponding marking.
Even the exchange of individual words can significantly complicate the search. In a test with English texts, the program initially recognized about 92 percent of the marked content. After ten percent of the words were replaced with synonyms, the hit rate dropped to 66 percent.
Short texts are also more difficult to reliably examine. A positive hit does not prove that the entire content originated from an AI. The marking can remain even if a human subsequently edited the text.
The watermark neither shows the extent of human involvement nor the accuracy of the statements contained. It also cannot determine who used ChatGPT for the creation or editing of the text.