On by default, across seven models
Text coming out of the Grok API carries an invisible watermark from 7 November. SpaceXSI, the Elon Musk company that builds the Grok models and whose legal entity is SpaceX Super Intelligence LLC, documented the change as covering seven models — grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants — and says it is on by default for all of them.
The scope is wide. The documentation says watermarking is enabled on every one of the company’s API endpoints, the global api.x.ai and any regional endpoint, and that it applies to every text-generation API, including the Responses API and the Chat Completions API. It covers ordinary replies, structured outputs and the arguments of tool calls. There is no request parameter to set, and the only change to the response shape is a new x-watermark header a developer can read to check whether a particular output was marked.
The company gives one reason for doing it: “SpaceXSI watermarks text model output to comply with the EU AI Act.” It names no article.
What the rule actually says
The obligation it is answering is in Article 50(2) of the AI Act, which requires providers of AI systems “generating synthetic audio, image, video or text content” to ensure their outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated”, with technical solutions that are “effective, interoperable, robust and reliable as far as this is technically feasible”.
That wording is why the mark is statistical rather than typographic. SpaceXSI says the watermark is applied while the model generates text, slightly shaping which tokens it picks, so it adds no characters, markup or hidden text — and therefore survives being copied out of a chat window and pasted somewhere else. Detection is probabilistic too: the company scores a passage and reports it as watermarked, not watermarked or uncertain, and says longer text gives a stronger signal.
Who can run that check is a separate matter. Access to the detection API is granted by application, through a form, so the ability to mark text is now default and universal while the ability to read the mark is not.
Turning it off is a legal statement
The opt-out is the part worth reading closely. Watermarking is set per model for an entire team, and only team admins see the control, in a Console row that reads On (all models), Off (all models) or Custom. The setting applies to every API key on the team, and changes take effect on new requests within five minutes.
Switching a model off is not a plain toggle. The documentation says opting out “asks you to confirm that an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure, or other legal requirement”. Every opt-in and opt-out is written to the team’s audit log, where admins can search it by event ID, description or user. Some models can also be locked on for a team: they appear in the dialog without a checkbox and cannot be turned off from the Console at all.
Taken together, that is a design choice as much as a compliance one. A developer who wants unmarked Grok text has to be an admin, has to assert in the product that the law exempts them, has to accept responsibility for any disclosure obligation themselves, and leaves a dated record of having done so. The things to watch next are whether the detection API opens beyond applications, and whether the consumer Grok apps — which the documentation does not cover — follow the API.
