Text returned by the Grok API will carry an invisible watermark by default from November 7, 2026. xAI states in its developer documentation, “Output Watermarking,” that the change covers seven Grok models and is intended to comply with the EU AI Act. Developers do not need to change their request format, but teams should confirm their watermark settings before November 7, or their output will be marked by default.
All seven Grok models are covered, including tool-call arguments
The seven models listed in the documentation are grok-4.7, grok-4.6, grok-4.5, grok-4.3, and the reasoning, non-reasoning, and multi-agent versions of grok-4.20. The watermark applies to all of xAI’s API endpoints, including the global endpoint and regional endpoints, and to requests through both the Responses API and the Chat Completions API. The mark is not limited to ordinary prose: structured outputs and tool-call arguments generated by the model are marked too, which matters most for teams that wire Grok into automated workflows.
The watermark lives in token-choice probabilities, so short text is no proof
The watermark does not add characters, labels, or hidden text. Instead, it slightly steers the model’s token choices during generation, leaving a statistically detectable pattern. xAI says this does not affect response quality or speed, although no independent testing supports that claim yet. Whether a single response was marked can be checked in the x-watermark response header: a value of true means it was watermarked. Checking a passage requires a separate application for detection access, and results come back as watermarked, not watermarked, or uncertain; longer text gives a stronger signal, and rewriting or shortening weakens it.
For users, the change works on two levels: people reading Grok output will see no difference, while the teams running the API gain a compliance responsibility. Team administrators can turn watermarking off per model in the console, but before doing so they must confirm that a legal exception applies and accept responsibility for any marking and disclosure duties; the setting applies per team and model, not to an individual API key. Anthropic has already adopted a similar watermark for Claude text, and Grok’s move makes watermarking by default a more common choice on the API side. Before November 7, teams running content pipelines on the Grok API should check their console settings and add response-header checks to their workflows, rather than tracing the source of output only after it is questioned.
