xAI turns on text watermarking by default across the entire Grok API

xAI enabled invisible statistical text watermarking by default across all xAI API endpoints, covering text from seven Grok model variants. Developers can check whether a response carries a watermark through a per-request x-watermark header. Opting out is deliberately difficult: it requires team-admin approval for each model plus confirmation of a legal exception.

The driver is regulation. Article 50(2) of the EU AI Act requires providers of generative AI systems to mark synthetic output in a machine-readable, detectable way, with obligations applying from December 2. Statistical watermarking subtly biases token choices to leave a pattern that a detector can identify without visibly changing the text.

xAI's default-on approach is more aggressive than many rivals have publicly committed to for text. Google has deployed SynthID text watermarking in some products, but text watermarking has generally been treated as fragile. Expect other labs serving EU customers to announce their own compliance approaches in the next three weeks.

The caveats are well known. Text watermarks can weaken under paraphrasing or translation, and short outputs carry little signal. Developers building on Grok should check whether downstream editing breaks detection and whether watermarks affect output quality. In separate news, xAI launched the Grok Voice 3.0 API, and Musk said a C/C++ rewrite delayed Grok 4.8.