Elon Musk's company xAI has begun embedding an invisible watermark in texts generated by its Grok language models. Since November 7, 2026, responses from the grok-4.7, grok-4.6, grok-4.5, grok-4.3 models and three versions of the grok-4.20 model have carried it. The mark is added automatically for all developers who integrate Grok into their applications through an application programming interface (API). People cannot detect the watermark, but xAI can later find it and confirm that the text was written by its model. The company states in its documentation that this enables it to comply with the requirements of the European Artificial Intelligence Act.
The mark is hidden in the words
The watermark does not look like a logo in the corner of a photo. It does not add any characters, marks, or hidden code to the text. It is created as the model generates the text and subtly influences its choice of words. Simply put, for each subsequent word, the model slightly favors certain options over others. No one will notice this in a single sentence, but in a longer text these minor shifts combine to form a pattern that specialized software can detect.
The detector compares the text with the pattern and classifies it as watermarked, unwatermarked, or uncertain. The longer the text, the stronger the signal the software receives. However, xAI does not give everyone access to the detection tool; interested parties must apply for it. The mark applies to ordinary responses as well as structured data and inputs that the model passes to other tools. It is used on all xAI servers, not only for developers in Europe.
It can be disabled, but only with a declaration
Developers do not need to rewrite anything. Each response contains the technical header x-watermark, and the value true means that the text has been watermarked. Anyone wishing to remove the mark will encounter several obstacles. It can only be disabled by a team administrator in the control panel, separately for each model and simultaneously for all of the team's keys. In doing so, the administrator must confirm that a legal exemption applies to this use and that they assume full responsibility for any required labeling. Each such change is recorded in the activity log, and the watermark cannot be disabled at all for some models.
What this means for users
Anyone using an application built on Grok will not see any difference. The text looks exactly the same. According to xAI, the watermark does not reduce response quality or generation speed, although independent tests have not yet confirmed this. The difference only becomes apparent when someone wants to verify where the text came from.
The watermark also has its weaknesses. It cannot prevent misuse on its own, and statistical marks become weaker when someone rephrases the text or translates it into another language. Developers should also ensure that their systems do not delete the x-watermark header when saving data. Otherwise, they will lose the simplest evidence that the text from Grok contained the mark.
Competitors have so far proceeded more cautiously. Anthropic announced a second round of watermark testing beginning November 16, but only as a voluntary experiment. Other AI providers operating in Europe have three weeks to make a decision. For systems launched before August 2, 2026, providers must introduce technical labeling no later than December 2, 2026.
Source: docs.x.ai
