Grok API Watermark Is Now On by Default: What Developers Should Do

The Grok API watermark is now switched on by default. Since November 7, text from seven Grok models carries an invisible statistical mark, and only a team admin can turn it off.

If you build on Grok, your app keeps working with no code changes. But your users’ text is now traceable, and you should decide whether that is fine for your product.

Here are the quick facts:

• Start date: November 7, 2026, on by default.
• Covers seven models, from grok-4.20 up to grok-4.7.
• Applies to the Responses and Chat Completions APIs.
• Opt-out is per model, set by a team admin in the Console.
• Check the x-watermark response header to confirm it is active.

Those five points are enough for most developers. The rest of this post covers the details and the catches.

What exactly is the Grok API watermark?

It is a hidden pattern in word choice. During generation, the model slightly nudges which tokens it picks. Nothing is added to the text, so you won’t see extra characters or markup.

According to xAI’s developer documentation, the company says it marks text output “to comply with the EU AI Act.” xAI also says the mark doesn’t change response format, quality, or speed.

Detection is statistical. A detector sorts text into watermarked, not watermarked, or uncertain. Access to the detector is granted per application through a request form.

Which Grok models and outputs are covered?

Seven models are on the list. The mark applies to text from regular requests, structured outputs, and tool call arguments.

The last row matters. xAI gives no accuracy numbers, and short text almost certainly carries a weaker signal. Treat any detector result as a hint, not proof.

How to check whether the watermark is active

Look at the x-watermark response header on your API calls. A value of true means that output was watermarked. Any other value, or no header, means it wasn’t.

This is the quickest test. Log the header in staging for a day and you’ll know which of your models and keys are affected.

How to opt out of the Grok API watermark

Only team admins can change it. In the xAI Console, open Team settings and find the “Model output watermarking” row. It shows On (all models), Off (all models), or Custom.

Click the settings icon, clear the box for each model you want to exclude, save, and confirm with “Opt Out.” The change covers every API key on the team and takes effect for new requests within five minutes. Every change is recorded in the team audit log.

There is a legal catch. To opt out, you must confirm that an exception under applicable law covers your use, and that you alone are responsible for any marking or disclosure. Some models may also be locked on, which means you’d have to contact your admin or xAI support.

Why this matters beyond Grok

Text watermarking is becoming normal. Anthropic announced in August that it would mark Claude text, as City AM reported. xAI had not signed the EU’s transparency code at the time, so this default is a notable shift.

If you’ve read our guide on how Claude marks its text, the idea is the same. The difference is control: Grok gives admins a clear switch, though with a legal confirmation attached. Admins who also manage Claude may want our enterprise controls checklist.

Two practical risks are worth a thought. Ghostwriting or customer-facing writing tools may not want machine-detectable text. And a mark is not a guarantee, since heavy editing or translation may weaken it.

A simple plan for this week

You don’t need a big project. Three small steps cover it.

First, log the x-watermark header on a test call for each Grok model you use. Second, ask your legal or compliance contact whether marked text is a problem for your product. Third, if you do opt out, write down who approved it, since the Console confirmation makes your team solely responsible.

Also tell your support and content teams. If someone later runs a detector on text from your app, they should know the mark is expected, not a bug.

Frequently Asked Questions

Will the Grok API watermark change my output quality?

xAI says no. It states the mark doesn’t affect quality or speed, and nothing visible is added to the text.

Can I turn the watermark off for one model only?

Yes. The Console lets admins clear the checkbox per model. The setting then applies to every API key on the team.

Does the watermark apply to images or code?

The documentation covers text output only. Images, video, audio, and code aren’t mentioned, so don’t assume they’re marked.

Can anyone detect Grok watermarked text?

Not freely. Detection is statistical, and access is granted per application through a request form.

Should you leave it on?

For most apps, yes. It costs nothing, needs no code, and keeps you on the safe side of EU rules. Opt out only if your product truly needs untraceable text, and only after your legal team agrees the exception applies, because that confirmation puts the responsibility on you.