Grok Text Watermark: What Changes on November 7

Yes, Grok is starting to watermark text. From November 7, 2026, text that seven Grok models return through the xAI API carries an invisible watermark by default, on every endpoint and in every region, according to xAI's own developer documentation (xAI docs). The mark hides in word choices, not in hidden characters. You can see whether a response was marked by reading an x-watermark response header, team admins can opt out model by model, and only approved applicants can run the detector. The docs cover the API only; they say nothing about text typed into the Grok app or on X. Images and videos made in the Grok apps already carry a visible Grok watermark.

This guide covers what the Grok text watermark includes, how to check a response, what opting out asks of you, and how Grok's approach compares with Claude and ChatGPT.

What the Grok text watermark covers

xAI's documentation, which now appears under the SpaceXSI name, gives the reason in one line: the company watermarks text output to comply with the EU AI Act. The page is dated October 31, 2026, and lists these models for the November 7 start:

• grok-4.7
• grok-4.6
• grok-4.5
• grok-4.3
• grok-4.20-reasoning
• grok-4.20-non-reasoning
• grok-4.20-multi-agent

The three 4.20 variants count as separate models. Four details matter more than the list itself.

It is global, not EU only. The watermark is on at the global endpoint, api.x.ai, and at every regional endpoint. A developer in Texas gets the same marked output as one in Berlin.

It covers every text generation API. That includes the Responses API and the older Chat Completions API.

It covers more than prose. The docs name ordinary text responses, structured outputs and tool call arguments. If your app asks Grok for JSON or lets it call functions, those generated strings are in scope too.

Nothing changes in your code. There is no request parameter, and the response format stays the same apart from one new header.

How to check if a Grok response was watermarked

Every response from the Responses API or Chat Completions API can carry an x-watermark header. The docs define it narrowly: x-watermark: true means the output for that request was generated with the watermark. Any other value, or no header at all, means it was not.

In Python with the OpenAI SDK pointed at xAI, the docs show a raw response call like this:

raw = client.responses.with_raw_response.create(
    model="grok-4.7",
    input="Write a haiku about the ocean.",
)
print(raw.headers.get("x-watermark"))

Two things to keep straight. First, the header is a receipt for your own request, not a detector. It tells the developer who made the call whether that output was marked. It cannot tell anyone else whether a pasted paragraph came from Grok. Second, if you log responses for compliance, log this header with them. It is the only record you get of which outputs carried the mark, and it is cheap to store.

We did not call the API for this guide (no xAI account was used), so the header behavior above is as documented, not as observed.

How the watermark works

The docs describe the mechanism in plain terms. The watermark is applied while the model generates text: it slightly shapes which tokens the model picks. So it adds no characters, markup or hidden text. A token is a word or a piece of a word, and at each step a model usually has several reasonable options. A secret key nudges which of those options gets picked, and over a long passage the pattern of picks becomes something a key holder can test for.

Detection is statistical. xAI says its detector scores a piece of text and reports one of three results: watermarked, not watermarked, or uncertain. Longer text gives a stronger signal.

That design has the same consequences as other sampling watermarks such as Google's SynthID Text, described in a 2024 Nature paper (Nature):

• Short replies are weak evidence. A one line answer offers too few choices to test.
• Exact answers carry little mark. A number, a code token or a fixed field name leaves the model almost no choice.
• Rewriting erodes it. Heavy paraphrasing or a pass through another model changes the very choices the mark depends on.

xAI does not say which algorithm it uses, how strong the setting is, or whether low entropy tokens are skipped. It states that watermarking does not change response quality or speed.

Who can detect Grok text?

Not the public. Access to xAI's watermark detection API is granted by application through a form linked from the docs. The page does not say who qualifies. Anthropic, by contrast, names its eligible groups (regulators, law enforcement, media, fact checkers, independent researchers, educational organizations, EU civil society groups and enterprises with their own compliance duties) (Claude Help Center).

So a teacher or editor who suspects an essay came from Grok has no watermark check to run today. What exists are classifier detectors such as Pangram and GPTZero, which judge writing style rather than reading a key. They are a different kind of test, with their own error rates; see are AI detectors accurate and our step by step method for detecting AI writing.

Each company's watermark uses its own key. A Grok detector cannot see a Claude mark, and the reverse is also true. A negative result from one checker says nothing about text from other models.

Opting out: what xAI asks of you

Grok is unusual in letting API customers switch the mark off. The process in the docs:

1. A team admin opens Team settings in the xAI console and finds the Model output watermarking row, which reads On (all models), Off (all models) or Custom. Only admins see it.
2. The admin opens the dialog and clears the checkbox for each model to opt out.
3. The admin saves and confirms with Opt Out.

Changes reach new requests within 5 minutes, and opting back in works the same way. The setting belongs to the team, so it applies to every API key on that team.

The catch is legal, not technical. Opting out asks you to confirm that an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure or other legal requirement. Every opt in and opt out lands in the team audit log. Some models may be locked on for a team; they appear in the dialog without a checkbox and cannot be turned off from the console.

Our reading: treat the opt out as a compliance decision with a paper trail, not a settings tweak. If your product serves EU users and you are not sure an exception applies, leave it on and get advice first.

Grok vs Claude vs ChatGPT text watermarks

All three companies cite the same law. Article 50 of the EU AI Act has applied since August 2, 2026, and requires generative AI providers to mark output in a machine readable way so that it can be detected as AI generated (European Commission). Their choices differ:

Anthropic applies its mark worldwide because it says it lacks a durable way to scope it by region (Anthropic). Its help center table now lists text watermarks on older models such as Claude Opus 4.7 and Opus 4.6 on Anthropic's own surfaces. OpenAI's textGrain marks ChatGPT text in the EU and stays off in the API unless a developer enables it (OpenAI Help Center). Our ChatGPT watermark guide covers textGrain and the Claude rollout in detail.

The practical difference: Grok is the only one of the three that both marks API output by default and gives developers a documented per response flag plus a self serve switch to turn it off.

Images and video: the visible Grok watermark

Much of the search interest in a "Grok watermark" is about pictures, not prose. That is a separate system. xAI's consumer FAQ says images and videos generated on the Grok website and apps include a visible Grok watermark to show the content was made with AI, that there is no setting to remove it, and that removing, altering or obscuring the watermark or other provenance signals breaks its Acceptable Use Policy (Grok FAQ).

We found no xAI documentation of Content Credentials (C2PA) on Grok images. So the usual file checkers, such as Content Credentials or Google's SynthID Detector, are unlikely to report anything for a Grok image. Our SynthID Detector guide explains which companies that tool can read.

What outside research says about text watermarks

No one outside xAI can test its deployed watermark. The closest independent evidence is a September 2026 paper from Case Western Reserve University that tested the open source SynthID Text implementation on two open weight models, Gemma-2-9B and Llama-3.1-8B (arXiv). Its findings are useful as a guide to the method, not to Grok specifically:

• Prose quality: the measured effect of the watermark did not exceed the effect of changing the random seed.
• Code quality: correctness fell 3.1 points on one model (60.7 percent vs 63.8 percent) and showed no measurable change on the other.
• Detection on prose: at a 1 percent false positive rate, the detector caught the mark in 39 percent of 200 token passages and 56 to 59 percent of 400 token passages.
• Detection on code: close to chance, with an AUROC of 0.55 and 0.57, where 0.5 means a coin flip.

That last point applies directly to Grok, since xAI marks structured outputs and tool call arguments, both low choice formats. Expect the mark to be faint there. The authors' wider argument is that vendors' quality and privacy claims cannot be verified from outside, and they call for released test outputs, disclosed configurations, audits, a shared evaluation protocol and a common detection interface. For marks that identify a model rather than a piece of content, see AI model fingerprinting. For how watermarks fit next to classifiers and provenance metadata, see what is AI detection.

What this means for you

• Grok API developers: nothing breaks on November 7. Log the x-watermark header, check which models are locked on for your team, and decide on opt outs with whoever owns compliance.
• Apps that pass Grok text to users: assume EU readers may later have it checked by an approved detector. Long, freely worded answers carry the strongest mark.
• Teachers and editors: there is still no public Grok text checker. Do not treat a "watermark scan" website as evidence.
• Grok app users: xAI's docs do not say consumer chat text is marked. Images and videos from the app carry the visible watermark either way.

FAQ

Does Grok watermark text?

Yes, through the API. From November 7, 2026, text from grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the three grok-4.20 variants carries an invisible watermark by default. xAI's docs do not say whether text in the Grok app or on X is marked.

Can I check whether text came from Grok?

Not with a public tool. xAI's detection API is available only by application. Developers can see whether their own responses were marked by reading the x-watermark header, but that works only for requests they made.

Can I turn off the Grok watermark?

Team admins can opt out per model in the xAI console, with changes applied within 5 minutes. Opting out requires confirming that a legal exception applies and that you take responsibility for any marking duties. Some models may be locked on.

Does copying Grok text remove the watermark?

No. The mark lives in the word choices, so it travels with copied text. Heavy rewriting, translation or paraphrasing by another model can weaken it, and short passages may be too brief to test.

Does the Grok watermark change answer quality?

xAI says it does not affect quality or speed. Independent tests of a similar open source method found no measurable effect on prose and a small cost on code for one of two models, but no outside group can test Grok's own setup.

Why do Grok images have a watermark?

Images and videos made on the Grok website and apps include a visible Grok watermark to show they are AI generated. There is no setting to remove it, and xAI's Acceptable Use Policy prohibits removing or obscuring it.