Starting today, November 7, 2026, text generated through xAI’s API carries a signature you cannot see. The company, which now calls its AI unit SpaceXSI, has switched on invisible watermarking by default for seven Grok models, so machine-generated text can be identified by software even though it looks unchanged to anyone reading it.
The change is API-only and it is already live, which means developers building on Grok need to understand what is marked, how to check for it, and what happens if they turn it off. Here is everything that matters.
Which Grok models are watermarked
The watermark applies to seven models from November 7, 2026: 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. It is on by default, with a per-model opt-out for team admins.
The scope is broad. Watermarking covers every text generation API, including the Responses API and the Chat Completions API, on every SpaceXSI endpoint — the global api.x.ai host and the regional endpoints alike. It applies to ordinary text responses, structured outputs, and even tool-call arguments, so there is no output format that escapes the mark. If you are weighing how Grok compares with ChatGPT for your own stack, this watermark is now part of that comparison.
What the invisible watermark actually is
The watermark is not visible in the text itself. It adds no characters, no markup, and no hidden formatting. Instead it is a statistical pattern in the model’s word choices, applied while the text is being generated. A reader sees nothing different; a detector scores the text as watermarked, not watermarked, or uncertain, and longer passages give a stronger signal.
SpaceXSI says the watermark changes neither the quality of the output nor the speed at which it is produced. That is the vendor’s claim and it has not been independently measured, so treat it as a promise to verify in your own latency and quality monitoring rather than a guarantee.
How to tell if a response was watermarked
There is one visible technical change in your integration: a new response header. Read the x-watermark header on the Responses API or Chat Completions API reply. A value of x-watermark: true means that request’s output was generated with the watermark. Any other value, or a missing header, means it was not.
Full detection is not public. SpaceXSI limits its detection tools to approved organisations such as regulators, researchers, media, and fact-checkers, and access to the detection API is by application. For most developers, the response header is the practical check.
An independent explainer walks through the header check and the opt-out flow in detail.
Why xAI is switching it on now
SpaceXSI gives a one-line reason in its documentation: it watermarks text to comply with the EU AI Act. Article 50 of the Act requires providers of AI systems that generate synthetic text, audio, images, or video to ensure outputs are marked in a machine-readable format and detectable as artificially generated.
The timing is not accidental. The Act’s Article 50 transparency obligations became enforceable on August 2, 2026, with a grace period for systems already on the market running through December 2, 2026. Turning the watermark on by default a few weeks before that deadline puts SpaceXSI on the compliant side of the line — and shifts the compliance burden onto any customer who switches it off.
How to opt out — and what it costs
Team admins can turn watermarking off for some or all models in the SpaceXSI Console, under team settings. The toggle takes effect within about five minutes and every change is logged in the team audit trail.
Opting out is not a free choice, though. To disable it, a team admin must confirm that a legal exception applies to their use and accept sole responsibility for any remaining marking or disclosure obligations. Some models may also be locked on. In short, the default is compliance; opting out means asserting you have a legal reason not to comply and owning the consequences.
What this means for developers building on Grok
The docs scope the change to API text generation, and the docs say nothing about Grok on X, the grok.com chat interface, or the mobile apps — so do not assume consumer Grok carries the same mark. What is covered is the Grok API as developers actually use it today: agents, SaaS features, search tools, and internal copilots. As enterprise AI spending climbs and agents take over more workflows, more of that output will carry the mark by default.
That has practical consequences. Text your app generates through the API keeps its statistical signature when it is copied, pasted, or published elsewhere, including into a blog or CMS. If your product generates content for users in the EU, the watermark may already satisfy part of your disclosure duty — but if you opt out, that duty falls squarely back on you. Cursor’s Voice Chat with Agents runs on Grok Voice, which shows how deeply Grok is now embedded in developer tooling, and every one of those integrations inherits the new default.
Watermarking is becoming the industry default
Grok is not alone. Anthropic’s Claude now adds an invisible watermark to the text it writes, with models launched on or after August 2, 2026 watermarking from day one, and the mark extended to four more models on November 5. Google marks its AI outputs too. With Anthropic, Google, and now xAI all marking text, a watermark signal is fast becoming a standard part of an API response rather than an exception.
The November 7 AI brief confirms the switch took effect today alongside OpenRouter’s routing rebuild.
Expect the tooling around it to mature quickly: headers today, detectors and disclosure helpers tomorrow, and eventually compliance checks built into the CI pipelines of any team shipping AI-generated content.
What to do this week
- Check for the x-watermark header in your Grok API integration and log it alongside your existing response metadata.
- Review your team’s watermarking settings in the SpaceXSI Console and confirm which models are on, off, or custom.
- Decide your opt-out position with legal advice, not on a developer’s hunch — opting out means accepting sole responsibility for disclosure.
- Test structured outputs and tool-call arguments specifically, since those are covered too and are easy to overlook.
- Update your product documentation and user-facing disclosures to reflect that API-generated text is watermarked by default.
Invisible watermarking will not change what Grok writes, and most users will never notice it. But it changes who is responsible for saying that text is AI-generated, and it makes that answer visible to machines even when it is invisible to people. Developers who check their headers, review their console settings, and update their disclosures this week will be ahead of the December 2 deadline — and ahead of the next provider to flip the same switch. For a broader look at the tools shaping this shift, see our roundup of the best AI tools of 2026.
