xAI’s official documentation says text watermarking will apply from November 7 to designated Grok API models, including versions 4.7, 4.6 and 4.5. It adds a statistical pattern through word selection in normal text, structured responses and tool-call arguments rather than hidden characters or formatting changes. An x-watermark response header set to true confirms that the particular output was marked. Administrators can request model-level exceptions for regulatory or legal needs, although some models may not allow disabling it. Detection is probabilistic and works better on longer text, so operators should distinguish checking the response header from later statistical detection.