Grok API Now Watermarks Text by Default

Flat illustration in cobalt blue, warm yellow and white showing the Grok API watermarking AI-generated text by default.

 

From 7 November 2026, text generated through the Grok API carries an invisible watermark by default. If you build apps, tools or content workflows on top of Grok, the text your users receive is now marked as AI-generated — and you did not have to change a line of code.

The change covers seven Grok models: 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 applies to text, structured outputs and tool-call arguments, across both the Responses API and the Chat Completions API, at every endpoint including the global api.x.ai endpoint.

Why xAI made the change

The reason stated by xAI is compliance with the EU AI Act. Article 50 of the Act requires providers of generative AI systems to mark their outputs in a machine-readable and detectable way, so that people and systems can tell AI-generated content apart from human-created content. The marking deadline is 2 December 2026, so the Grok change lands just under a month before it.

This is not a product feature or a marketing move. It is a legal requirement that applies to every major provider of generative AI, and other providers are rolling out similar measures in the same window. Anthropic, for example, began watermarking Claude text in early November too, with a second wave of models to follow.

What changed for developers

The good news for developers is that almost nothing changes in how you call the API. No code change is needed, and the response format stays the same, with one addition: a new x-watermark response header. When it reads true, the output has been watermarked.

The watermark itself is invisible. It adds no characters and no hidden text to the response. What you see is what you get — the same words, the same structure. xAI says the watermark does not affect the quality or speed of the output, although that claim has not been independently tested.

That matters because watermarking has a history of trade-offs. Some early approaches changed word choices in ways that made text read slightly worse, or slowed generation down. xAI's position is that its approach avoids those costs, but developers will only know for sure once they run the watermarked models at scale and compare results with what came before.

Who can opt out — and who cannot

Team admins can switch the watermark off, but only per model, and only in the SpaceXSI Console, under Team settings and the "Model output watermarking" option. SpaceXSI is the name used for the xAI unit in the developer documentation.

Opting out is not a casual toggle. An admin must confirm that a legal exception applies and accept sole responsibility for marking the output. In other words, turning the watermark off does not remove the obligation — it moves the responsibility onto the team doing it. Some models are locked on, meaning the watermark cannot be switched off for them at all.

This is a sensible design for a compliance feature. The default protects everyone who does nothing, and the escape hatch exists only for teams that have a genuine legal reason and are willing to own the consequences.

How detection works

A watermark is only useful if someone can check for it. xAI provides a detection API that reports whether a piece of text is watermarked, not watermarked, or uncertain. Longer text gives a stronger signal, which makes sense: the more text there is to examine, the more confident the check can be.

There is a catch. Access to the detector is limited and not public. Regular users cannot check a piece of text themselves. The same limitation applies on Anthropic's side, where the detection API is in private preview for eligible organisations such as regulators, media outlets and researchers.

That means the watermark is mainly a tool for platforms, regulators and compliance teams, not something an ordinary reader can use to verify a suspicious paragraph. It still matters, because it gives those organisations a reliable way to confirm where text came from.

What is not covered

It is important to be clear about the scope. The watermark applies to the API only. The Grok app, the Grok website and Grok on X are not covered by this change. Grok Imagine, the image and video tool, has a separate, older watermark of its own.

So if you are a developer using the API, your outputs are marked. If you are a user chatting with Grok in the app, this particular change does not apply to you — at least not yet.

Key takeaways

  • From 7 Nov 2026, text from seven Grok models via the API is watermarked by default.
  • The change is driven by EU AI Act Article 50, ahead of the 2 Dec 2026 marking deadline.
  • No code change is needed; a new x-watermark header shows whether output is watermarked.
  • Team admins can opt out per model in the SpaceXSI Console, but must confirm a legal exception applies and accept sole responsibility.
  • Detection is through xAI's detection API, but access is limited and not public.
  • The Grok app, website and Grok on X are not covered; this is an API-only change.

What this means for you

If you run a product on the Grok API, the practical impact is small: your outputs are now marked, your code does not need to change, and you can check the x-watermark header if you want to confirm. If you are in a team that is considering opting out, think carefully. The opt-out exists for genuine legal exceptions, and it puts the marking responsibility squarely on you.

For everyone else, this is a sign of where the industry is heading. By early December, watermarked AI text will be the default across the major providers, not the exception. The interesting question is no longer whether AI text gets marked, but whether the marking systems work well enough — and are open enough — for the people who need to rely on them.

Conclusion

Grok's watermark is a quiet, technical change with a loud legal reason behind it. It does not change what the models write, and it does not ask developers to do anything new. But it does mark a shift: AI-generated text is now labelled at the source, by default, and the days of unmarked machine-written content from the big providers are coming to an end.

FAQ

Does the Grok watermark change the text I receive?

No visible change. The watermark adds no characters or hidden text, and the response format stays the same apart from a new x-watermark header. xAI says it does not affect quality or speed, though that has not been independently tested.

Can I turn the Grok watermark off?

Team admins can opt out per model in the SpaceXSI Console, but they must confirm a legal exception applies and accept sole responsibility for marking. Some models are locked on and cannot be opted out.

Is the Grok app watermarked too?

No. This change covers the API only. The Grok app, website and Grok on X are not covered, and Grok Imagine image and video outputs have a separate, older watermark.

Sources

Related reading on Daily Tech & AI:

Researched with AI assistance and reviewed for accuracy.

Follow on WhatsApp

Follow Daily Tech & AI on WhatsApp for the latest tech and AI updates: https://www.whatsapp.com/channel/0029Vb9XzdHDJ6H4TFNAl61r

Comments