xAI will watermark Grok API text by default from November 7, with an admin opt-out

xAI has updated its developer documentation to say that, starting November 7, 2026, text generated through the Grok API will carry an invisible watermark by default. The change, reported by Superpower Daily on October 31, applies on the global endpoint and on regional endpoints, and the documentation presents it as a step to comply with the EU AI Act.

Which models and outputs

Seven models are listed: grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning and grok-4.20-multi-agent. The mark covers text produced by both the Responses API and the Chat Completions API, including plain text, structured outputs and tool-call arguments. The documentation does not say how streamed responses are handled.

The technique is statistical. Rather than inserting characters, markup or hidden text, it slightly influences which tokens the model picks during generation. Detection returns one of three verdicts (watermarked, not watermarked or uncertain), and longer passages give a stronger signal. Access to the detection API is granted application by application through a request form. xAI says the watermark does not affect quality or speed.

Developers do not need to change their code. The only visible difference is a new x-watermark response header: a value of true means the output was marked, while any other value, or no header at all, means it was not.

An opt-out with conditions

The setting can be changed only by team administrators, in the Console's team settings, model by model, and it applies to every API key on the team. Turning it off requires confirming that a legal exception covers the use case and accepting sole responsibility for any related obligations. Changes take effect for new requests within five minutes and are recorded in the team audit log. The documentation warns that some models may be locked on for a given team.

Context

The move brings xAI closer to its competitors. In August, City AM reported that xAI had not signed the EU's code of practice on transparency for AI-generated content, while still being bound by the AI Act's requirements. Anthropic already embeds a similar invisible mark in Claude's text, which it does not let customers disable, and is extending it to older models, as this site reported on November 1. Because xAI's signal is probabilistic and teams with a legal exception can switch it off, a passage that tests negative says little on its own; the documentation also gives no figures on detection accuracy.