Seven Grok models are scheduled to add an invisible watermark to text from the API by default from November 7. The plan covers ordinary replies, structured output and tool-call arguments. Team administrators will manage eligible exceptions by model. The mechanism changes token selection during generation, while detecting its statistical pattern requires access to the company’s separate detection service.
SpaceXSI, the company behind Grok, plans to begin default watermarking of text generated through its application programming interface, or API, on November 7. The schedule covers Grok 4.7, 4.6, 4.5 and 4.3, plus the reasoning, non-reasoning and multi-agent versions of Grok 4.20. Ordinary replies, structured output and the arguments supplied to tool calls fall within the stated scope, across global and regional endpoints.[1], [2]
The watermark is applied by slightly influencing the tokens, or small pieces of text, chosen during generation. According to SpaceXSI, no visible marker or extra hidden text is inserted. Applications can read the x-watermark response header: true identifies a reply generated with the mark. Detection of copied text is a separate statistical process that can return watermarked, unmarked or uncertain; the company grants detector access by application.[1]
Only team administrators can change the model-specific setting, which applies to every API key belonging to that team. Opting out requires confirmation that a legal exception covers the use and that the customer takes responsibility for applicable marking or disclosure duties. Some models can remain locked on. Changes are recorded in the team audit log and are expected to affect new requests within five minutes.[1]
