news

xAI's default Grok watermark covers text output only, not image or video generation

The documentation that went into effect on November 7 names seven Grok text models and defines the mark's scope as text generation requests.

By Priya Shenoy · 9 November 2026

xAI turned invisible output watermarking on by default on November 7, and the scope written into its own developer documentation is text. The page states that the company watermarks text model output to comply with the EU AI Act, and that marking applies to text output from text generation requests for each model that has it enabled. The generators that produce pictures and clips are not part of that policy, which is a notable gap given that the transparency rule being cited is aimed squarely at synthetic media.

Seven text models, every endpoint, one header

The documentation, published under the SpaceX Super Intelligence API branding that now appears across xAI's developer pages, says watermarking is enabled on every endpoint, including the global api.x.ai endpoint and regional ones, and that it covers output from every text generation API, both the Responses API and Chat Completions. Outlets reading the model list name seven: grok-4.7, grok-4.6, grok-4.5, grok-4.3, and the reasoning, non-reasoning and multi-agent variants of grok-4.20. Older models including Grok 4.1 Fast, Grok 4 and Grok 3 are not on it.

There is no request parameter to set. Control sits at the team level in the console, where an admin can switch watermarking off per model, and the only per-request evidence a developer gets is the response header, where x-watermark: true means that output was generated with the mark. The docs state the response format is otherwise unchanged. What the page does not describe is a public detector, so a third party holding a suspect block of Grok text has no documented way to check it; the header only helps the caller who made the request.

Article 50 is about synthetic media, and the marked models do not make any

The stated legal basis is the EU AI Act's transparency chapter, which asks providers of generative systems to mark synthetic output in a machine-readable format where technically feasible. Those obligations applied from August 2, 2026, with a transition running to December 2, 2026 for systems already on the market before that date, which is the deadline driving the current wave of provider announcements.

That framing makes the omission concrete rather than academic. xAI sells Grok Video and runs Grok Imagine as a consumer image and video product, and image and video are the content types the deepfake provisions of Article 50 were written for. The only marking scope xAI has documented for API output addresses text.

Marking matters more for video because detection does not work

Generation-time marking is load-bearing for video in a way it is not for text, because the after-the-fact alternative is weak. OpenAI and Meta told a United States evidence-rules panel this autumn that no method reliably detects AI video, which leaves provenance attached at generation as the only signal with a known origin. A statistical token-shaping watermark on an LLM response and a provenance signal embedded in a video file are different mechanisms with different failure modes, and the page published on November 7 only describes the first.

For teams building on the Grok API, the practical read is narrow and clear: text from the seven listed models is marked unless an admin has turned it off, the header is the receipt, and anything generated outside text generation requests falls outside what xAI has committed to in writing.

Sources: xAI, EU AI Act Article 50 guide.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.