The Pulse
xAI to Watermark Seven Grok API Models by Default on Nov. 7
xAI says invisible watermarks will be added by default to text from seven Grok API models starting November 7, 2026. Team administrators can opt out per model, while the API adds an x-watermark response header and requires no request-code c
AI.info Team · Written with AI · fact-checked ·
xAI plans to add invisible watermarks to text from seven Grok API models starting November 7, according to developer documentation posted October 31. The setting will be on by default across the company’s API endpoints, including its global endpoint, but team administrators can opt out model by model. The change applies to API text generation, not just requests routed through a particular regional service.
Seven Grok models get the default mark
The listed models are grok-4.7, grok-4.6, grok-4.5 and grok-4.3, along with grok-4.20 in reasoning, non-reasoning and multi-agent versions. The mark covers text returned through both the Responses API and Chat Completions API. It also applies to structured outputs and tool-call arguments, not only ordinary prose.
xAI says the watermark is intended to let the company later recognize text as model-generated and that it is adding the feature to comply with the EU AI Act. The documentation says the mark does not add visible characters, markup or hidden text to the response. Instead, it slightly shapes token selection as the model generates text.
Developers get a response header, not a new parameter
Applications do not need a new request parameter or a change to their integration. The response format remains the same apart from an x-watermark header on Responses and Chat Completions API replies. A value of true indicates that the output was generated with a watermark; a different value or no header means it was not.
The marking process itself is invisible, and checking for it is statistical. xAI says its detector can classify text as watermarked, not watermarked or uncertain, with longer samples providing a stronger signal. Access to the detection API requires an application to the company, so the documentation does not describe a public, self-serve detector for developers or readers.
Opting out shifts responsibility to the team
Team administrators can disable the setting for some or all models through the SpaceXSI Console, the name used in the documentation. The choice applies across every API key belonging to that team, and changes to new requests take effect within five minutes. Some models may be locked on and cannot be switched off through the console.
To opt out, administrators must confirm that an exception under applicable law covers their use and accept responsibility for any marking or disclosure requirements. The setting is therefore more than a technical toggle: it assigns compliance decisions to the API customer when the customer chooses not to use the provider’s watermark.
The scheduled start is November 7
The documented start date is November 7, a week after the page’s last update on October 31. Until then, the page describes the default as a forthcoming change; it does not say that all listed models are already returning watermarked text. The scope is limited to text generated through the API, and the documentation does not say whether the Grok consumer app or Grok on X will use the same watermark.
For teams using the API, the immediate implementation check is whether their systems preserve response headers and whether administrators want the default applied to each listed model. The output itself will look unchanged; the signal is carried in token choices and the response header, rather than a visible label attached to the text.