xAI to watermark Grok API text by default from November 7, including tool-call arguments

xAI is turning on text watermarking by default. From November 7, supported Grok API models, grok-4.7, 4.6, 4.5, 4.3 and 4.20, will embed an invisible statistical signal in their output. Developers do not need to change code, and opt-outs are limited.

Statistical text watermarks typically work by subtly biasing token selection with a secret key so a detector can later estimate whether text came from the model. Unusually, xAI applies it not only to prose but to structured responses and tool-call arguments. That breadth is what developers are questioning: biasing tokens inside JSON or function arguments could in theory affect determinism or exact-match outputs, although xAI implies quality is unaffected.

xAI frames the move as EU AI Act compliance, which requires machine-readable marking of synthetic content. Most labs have watermarked images and audio; default text watermarking across an entire API is rarer. Google's SynthID-Text is the best-known precedent, but default enforcement on tool calls goes further.

Open questions: how robust the watermark is to paraphrasing, whether detection tools will be public, and how enterprise customers can opt out. Developers on Dev.to and elsewhere want clarity on structured-output impact before the switch flips.