Watermarks will soon also mark AI texts from ChatGPT

Artificial intelligence recognize watermarks in AI texts - soon also at ChatGPT Status: October 22, 2026 • 5:30 a.m

Does this text come from an artificial intelligence? Watermarks that mark AI-generated texts can provide information. OpenAI has now announced this for ChatGPT.

A new EU AI Act law has been in effect since the beginning of August, requiring machine-readable labels for AI-generated content. AI providers such as Anthropic and Google have announced corresponding watermarks or are already using them in some cases.

Anyone who is online today most likely reads AI texts every day - without noticing it. Chatbots write product descriptions, applications, university term papers, even political statements. In the future, watermarks for AI texts will indicate whether a text comes from a machine.

When installing a watermark, a peculiarity of the language is used: there are often several wordings for the same content, for example "I'm going to the cinema today" instead of "I want to go to the cinema today". In these linguistic freedoms, the AI ​​can be controlled unobtrusively when generating text.

More and more presumably AI-generated children's books are appearing on Amazon. more

Normally the model calculates a probability for each possible next word and then "rolls" which one is chosen. For the watermark, this probability die is slightly "pointed": certain alternative words that do not change the meaning are given minimal preference and others are minimally disadvantaged. This creates a statistical pattern in the text that is not visible to readers.

This slight change is recalculated for each word, so the changes remain invisible throughout the text. The adaptation of the AI ​​texts cannot be recognized by both humans and machines, even after a lot of recalculation. Whether a text has such a watermark can only be recognized later if you know the secret "key" that describes how exactly the "cube" was manipulated.

Waldemar Berchtold, group leader for multimedia security at the Fraunhofer Institute for Secure Information Technology in Darmstadt, explains why this is the case: "If everyone can understand it, they can remove the watermark, they can undo it, so to speak, and they can also create a watermark."

Anthropic wants to add an invisible watermark to texts from its chatbot Claude. more

Berchtold points out that watermarks also have technical limitations. On the one hand, a text must have a certain length for the watermark technology to work. The researcher says that it only works well to hide and recognize watermarks when they are around 200 to 300 words long. On the other hand, there are problems when you change the text: "Because I also change the output and bring my style back in, and in doing so I basically obscure the trace of the AI ​​generator - the watermark."

The elegance of watermarks lies in the fact that the symbol is hidden in the text itself. However: If you change the text, you also change the watermark. Therefore, the decryption method can only provide a probability that the text contains a watermark, not 100% certainty.

As of today, there is a labeling requirement for AI content in the EU. But how secure are watermarks? more

Benno Stein, computer science professor at the Bauhaus University Weimar, points out that the AI ​​providers with watermarks ultimately do not just follow EU regulations. They also have a vested interest in being able to identify which texts come from the AI ​​and which do not. "If you are a language model manufacturer, then you have to constantly fill these language models with good texts," explains the researcher. "It is in the interest of language model manufacturers to recognize their own texts. This benefit is often overlooked."

It could soon become easier for consumers, but also institutions, to recognize AI-generated texts: In the future, AI providers will also offer detectors to be able to test texts for their origin. A probability statement will then probably be made such as: “This text was created by our AI model with a probability of 80 percent.”

This topic in the program: tagesschau.de | Watermarks for AI-generated texts | Oct 21, 2026 | 5:00 p.m