xAI Now Watermarks Grok API Text by Default. Here's What an Invisible Watermark Can and Can't Prove.

If you've used xAI's API in the past few days, your responses may have a hidden signature you can't see and can't copy-paste out.

xAI's documentation, last updated Oct. 31 and branded SpaceXSI on the docs site, says that starting Nov. 7, 2026, text from seven Grok models includes an invisible watermark by default. The page gives a plain reason: the company says it's doing it to comply with the EU AI Act.

Which Models

The default applies to 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 covers both the Responses API and Chat Completions, and it applies to ordinary text, structured outputs and tool-call arguments. Developers don't have to change their code.

How It Works

The watermark doesn't add characters, markup or hidden text. According to the docs, it slightly shapes which tokens the model picks while it generates, leaving a statistical pattern that can be recognized later. xAI says it doesn't change response quality or speed.

Developers can check whether a specific response was watermarked by reading an `x-watermark` header. A value of `true` means yes. Anything else, or no header, means no.

Detecting the pattern in a piece of text is a separate step, and the docs describe it as statistical. A detector scores text as watermarked, not watermarked or uncertain, and longer text gives a stronger signal. Access to the detection API is by application, through a form linked in the docs.

Opting Out

Team admins can turn watermarking off, per model, in the Console's team settings. The change applies to every API key on the team, takes effect within about five minutes, and is recorded in the team's audit log. Some models may be locked on and won't show a checkbox.

There's one catch. To opt out, an admin has to confirm that an exception under applicable law applies to their use, and that they're solely responsible for any marking or disclosure requirements.

Our Take

For publishers and platforms, the question that matters is what a watermark lets you conclude, and the answer is less than the word suggests.

It's a lock that only xAI can open. The detector is available by application, so there's no public way for a newsroom or a school to check a piece of text. A "watermarked" result is evidence that a Grok model probably wrote or helped write the text. A "not watermarked" result tells you much less: the text might come from a different model, from a team that opted out, from a model that isn't covered, or from a human. The absence of a watermark is not proof of human writing.

Statistical watermarks in general are also weaker on short passages, and can be diluted when text is heavily edited or paraphrased. xAI's page says longer text gives a stronger signal, but it doesn't say how the system performs on edited text or how often it gets things wrong.

None of that makes the move pointless. It's a compliance step, and one more signal in the pile. But if you're a publisher hoping to catch AI text with it, you can't, at least not yet.

What We Cannot Conclude

• We don't know the error rates. The docs don't give false-positive or false-negative figures.
• We don't know how it holds up after editing. Nothing on the page addresses paraphrased or rewritten text.
• We can't test it ourselves. Detection access is by application, and we haven't had it.
• We can't say how many customers will opt out. The setting exists, and the legal confirmation required to use it is a high bar.
• This covers xAI's API only. Other providers run their own systems, and we haven't compared them.

Why It Matters

• Regulation is showing up inside APIs. A legal requirement in the EU is changing default behavior for developers.
• Only some AI text can be flagged. Watermarks cover opted-in models from cooperating providers, not the whole internet.
• Publishers shouldn't treat a missing mark as proof. It's evidence at most.
• There's an audit trail. Opt-ins and opt-outs are logged, which matters if a regulator ever asks.

For the other dates and rules landing on AI developers this winter, see our calendar of AI events and deadlines.