xAI said on Oct. 31 that it will begin adding invisible text watermarks to supported Grok API outputs starting Nov. 7, 2026. The default-on change applies to seven Grok models and spans both Responses and Chat Completions, including plain text, structured output payloads, and tool-call arguments.
For API teams, the immediate signal is the x-watermark response header, which indicates whether a returned output was marked. xAI is also offering a separate detection API by application, but the docs note that detector results are probabilistic rather than absolute: they can come back as watermarked, not watermarked, or uncertain.
Admins can disable watermarking per model in the xAI console, so this is not just a model update but a provenance setting that may vary across teams. If your pipeline drops response headers before storage, moderation, or tool execution, you may lose the easiest audit trail for whether Grok-generated text carried a watermark in the first place.