The listed models are Grok 4.7, 4.6, 4.5, 4.3 and three 4.20 variants. The watermark slightly shapes token choices without adding hidden characters. Detection is statistical, may return uncertainty and requires access granted by application.
Administrators can opt out per model in the console, except for locked models, while accepting responsibility for applicable marking duties. The x-watermark header indicates whether a response was marked. The provider cites the EU AI Act; the documentation alone does not establish legal compliance or resistance to every form of text editing.