On November 14, Elon Musk wrote on X that Grok text watermarking is enabled only where required by law or requested by the user. SIExpo read the original post. Current API documentation also describes text watermarking as off by default. This materially updates our earlier coverage of a plan to enable it automatically from November 7.
Configuration, output and verification are different
According to the documentation, an administrator selects the setting per model for the team. An API response can carry an x-watermark header indicating whether that particular output was marked. The page describes a separate default-on regime for voice audio, with restrictions on opting out. A short statement about text should therefore not be applied automatically to every audio product.
The limits of our verification matter. SIExpo did not run paid API tests or independently audit every model and region. We checked a public statement and documentation, rather than every actual account configuration. Those sources do not establish that all text ever generated by Grok lacked a watermark, or that the earlier plan necessarily operated exactly as announced.
What invisible marking means
The documentation identifies SynthID Text as the technique used. Google DeepMind's material and the research paper in Nature explain an approach distinct from adding a visible label: the signal is associated with the statistical selection of generated elements. Searching copied text for a special character is consequently not a reliable way to establish whether it carries such a signal.
The research examines trade-offs among output quality, detectability and resilience to transformations under particular experimental conditions. It is not a universal guarantee of flawless identification for every short, edited or translated passage. Classifying the provenance of a passage is also separate from determining whether its factual claims are correct.
DeepMind presents SynthID as a provenance tool for several kinds of AI-generated content. That provides background, but does not independently establish another vendor's product configuration. For Grok, the company's own documentation and the state of the relevant team or service remain the pertinent operational references.
What the change means in practice
Our editorial interpretation is that developers should not build accountability around an assumption about a default setting. In an integration, retaining model version, configuration and service-provided provenance information separately from the final text can make later review more meaningful. Absence of a signal should not be promoted into proof of human authorship.
Consider an application that archives an answer but discards its response metadata. Months later, its operators may know what text was shown but not which configuration produced it. A documented workflow preserves that distinction. This is an editorial example of why provenance records matter, rather than a claim that the current feature solves every attribution problem.
We do not treat Musk's statement as a legal determination of users' or publishers' obligations. A provider's setting and applicable requirements are different questions. The verified news is the updated description of Grok's behavior and the need to stop presenting an earlier rollout plan as an unchanging current fact.
The cover is an AI-generated editorial illustration, not a documentary photograph.



