Grok API text now carries an invisible watermark by default on seven models

Story So Far

Text generated through xAI's Grok API now carries an invisible watermark by default. The change took effect on November 7, 2026, and covers seven current Grok models, according to the company's developer documentation. The documentation now appears under the name SpaceXSI, the SpaceX unit that builds Grok.

The seven models are grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three variants of grok-4.20 for reasoning, non-reasoning and multi-agent work. The mark applies to every text generation endpoint, including the global api.x.ai address and regional endpoints, and across both the Responses API and the Chat Completions API. It covers ordinary replies as well as structured outputs and the text a model generates as arguments for a tool call. Developers do not need to change their requests: there is no parameter to send, and the only visible trace is a new x-watermark response header, which reads true when a reply was marked.

Team administrators can switch the mark off model by model from the console's team settings, but opting out requires confirming that an exception under applicable law covers the use and accepting sole responsibility for any marking or disclosure the law requires. Changes take effect within about five minutes, apply to every API key on the team, and are recorded in the team's audit log. Some models may be locked on and cannot be switched off from the console.

The documentation itself is the evidence for the core of this story: it names the November 7 start, lists the seven models, and states plainly that the mark is on by default and that the reason is compliance with the EU AI Act. What the page does not do is name the article of the law or give a deadline; the December 2 date circulating in coverage comes from independent analysis of the Act's transition rules for systems already on the market, not from xAI. The page also says nothing about Grok's consumer apps or its presence on X, so the change as documented reaches developers using the API, not people chatting with Grok. And the company's statement that watermarking changes neither the quality nor the speed of responses is its own claim; no independent measurement of either has been published.

For teams building products on the Grok API, the practical effect is immediate: from November 7, the text their applications receive is marked as machine-generated unless an administrator has deliberately switched it off. That matters because the marking duty now sits with the provider by default, and a team that opts out takes the legal responsibility back onto itself. The trade-off is visibility: the mark is invisible and its detection statistical, with access to the detection tool granted only by application, so most developers can confirm a reply was marked only by reading the response header, not by checking the text itself.

The timing fits a wider pattern. Anthropic began watermarking text from four older Claude models on November 3, four days before Grok's switch-on, and OpenAI's equivalent mark stays off by default on its API, according to TechCrunch. The next date to watch is December 2, when the EU AI Act's transparency obligations reach systems that were already on the market, and whether xAI extends the same marking to Grok's consumer apps.