Grok Watermarks Text From Nov 7: What Creators Must Know

Starting November 7, 2026, text you generate through the Grok API will carry an invisible watermark by default. You will not see it, and your readers will not see it. But a detector built by SpaceXSI — the company formerly known as xAI, now a SpaceX unit — will be able to spot it, at least in longer pieces of text.

SpaceXSI published this in its own developer documentation, on a page titled “Output Watermarking,” last updated October 31, 2026. If you use the Grok API to draft blog posts, product copy, or client deliverables, this change touches your workflow directly. If you only use the Grok app or Grok on X, it does not — not yet. No date has been announced for those.

The short version: seven Grok models get watermarked text output starting November 7. The watermark is statistical, not visible. Developers get one new signal per response, an HTTP header called x-watermark. Detection itself is not public — access is by application only. And opting out is not a simple toggle: it is an admin-only, per-model decision that requires confirming a legal exception applies and accepting sole responsibility for marking and disclosure.

What exactly changes on November 7

From November 7, 2026, text generated by seven Grok models through the SpaceXSI API carries an invisible statistical watermark by default: 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.

The watermark applies to every text API — the Responses API and the Chat Completions API — at every endpoint, including the global api.x.ai endpoint and regional endpoints. The stated reason is compliance with the EU AI Act, but the rollout is not EU-only. A developer calling the API from the United States gets watermarked text too.

It also covers more than a plain chat reply. Ordinary text responses, structured outputs, and tool-call arguments are all watermarked. If the model produces text through the API, the marking applies.

What it does not cover is just as important. This is text only — not images or video from Grok Imagine, and not voice. And it is API only. SpaceXSI has announced no date for watermarking text in the Grok app, Grok on X, or the Grok website.

The legal basis is Article 50 of the EU AI Act, which requires providers of generative AI systems to mark synthetic content in a machine-readable format that can be detected as artificially generated. The transparency obligations have applied since August 2, 2026, and providers whose systems were already on the market before then have until December 2, 2026 to comply. November 7 falls 25 days before that deadline. SpaceXSI says watermarking changes neither quality nor speed — the company’s own claim, not independently verified.

How the watermark actually works

The watermark is statistical. When Grok generates text, it picks words — tokens, in model terms — one at a time from a ranked list of likely next words. A statistical watermark slightly shapes those choices, nudging the model toward certain words and away from others in a pattern that follows a secret key only SpaceXSI holds. The result reads like normal text, with no visible characters added and no hidden text tucked in — nothing a reader could spot by eye.

What changes is the distribution of word choices across a passage. Over a long enough stretch, that nudged pattern becomes detectable, if you have the detector and the key. Over a very short stretch, it may not be. Longer text carries a stronger, more detectable signal, while very short text may carry no detectable mark at all.

The one signal developers actually get

The only per-response signal developers get is a new HTTP response header: x-watermark. If it reads true, that response was watermarked. Any other value, or no header at all, means it was not. There is no score, no confidence percentage, and no per-sentence breakdown. You get a true-or-not flag in the headers, and that is the whole developer-facing signal.

That sounds thin, but it is also the most actionable part of the change, because it gives you something you can log. Recording that header value alongside each logged response is the single most useful habit to build this week.

Who can detect the watermark?

Ordinary readers cannot check a piece of text themselves. There is no public tool, no paste-a-paragraph-and-check website, and no browser extension from SpaceXSI for this. Detection access is by application only. As of November 6, 2026, it is still unanswered who exactly can get detector access, and under what terms.

What it does and doesn’t do

It marks text generated through the Grok API, by default, from November 7, for the seven listed models. It does so without visible changes to the text, gives developers a per-response header flag, and creates a statistical signal that a detector with the right key can identify in longer text.

It does not add visible characters or hidden text. It does not change the text your readers see. It does not mark images, video, or voice. It does not apply to the Grok app, Grok on X, or the website as of the announced date. And it does not, by itself, prove authorship in the way a signed document proves who signed it.

That last point is the most common misunderstanding about AI watermarking. A watermark is a statistical signal, not a certificate. Detection is statistical too, which means it has three possible outcomes, not two: watermarked, not watermarked, or uncertain.

A long, lightly edited Grok draft is likely to read as watermarked to a detector. A short reply may carry no detectable mark at all. A heavily edited passage — rewritten sentences, swapped paragraphs, translated into another language and back — may lose the signal or weaken it past the point of confidence. Anthropic says the same thing about its own marking: a mark “is not fully conclusive,” and heavily edited, translated, or very short text may carry none.

So the watermark does not settle, on its own, whether a given paragraph was written by a human or by Grok. It is one signal, strongest on long and lightly edited text, weakest or absent on short and heavily reworked text. And it does not solve the disclosure question for you: the watermark is machine-readable, your readers are humans, and a machine-readable mark does not replace a clear, human-readable disclosure where one is expected.

The opt-out trap

Opting out is possible, but it is not a toggle you flip in your own settings. It is a legal decision with a paper trail.

First, only team admins can opt out — not individual developers or API key holders. A team admin, in the SpaceXSI Console, per model.

Second, the admin has to confirm that a legal exception applies, and accept sole responsibility for marking and disclosure. The responsibility does not stay with SpaceXSI when you opt out. It moves to you.

Third, the opt-out applies to every API key on the team. Individual keys cannot be configured separately. You cannot watermark one project’s output and leave another project’s output unmarked under the same team.

Fourth, once set, the change takes effect within about five minutes. That is fast — good if you have made a careful decision, dangerous if you have not.

Put together, that is not a settings preference. It is a compliance call. If you administer the team yourself, you are the admin making that call; inside a larger team, it belongs to whoever owns legal and compliance.

Do not opt out casually, and do not opt out alone. If you think an exception might apply, document the reasoning, get the decision in writing, and keep that record with your API configuration notes. If you are not sure an exception applies, the default — watermarked output — is the safer place to stay. It is also still unanswered which legal exceptions SpaceXSI will accept, since the boundaries of that test are not public in the material I have verified.

Grok vs Claude vs ChatGPT text marking

All three big providers are now marking text, but not in the same way or with the same visibility. Here is the comparison, built only from what each company has actually said.

Grok is the most specific about which models are covered and what signal developers get; Claude is the most explicit about detection limits; ChatGPT’s marking is described in the least detail in the sources verified for this article.

What this means if you publish Grok-written text

If you use the Grok API to draft anything that ends up published — a blog post, product copy, a newsletter, a client deliverable — November 7 changes the facts on the ground, even if it does not change what your readers see. The text will look the same, but it will now carry a signal a detector can pick up, and you will have a header value you can log.

For blog posts, the watermark does not change your search ranking by itself, and it does not add a label to your page. What it changes is the audit trail. A long, lightly edited Grok draft is now detectable as AI-generated by anyone with detector access. That is not a reason to panic or to stop using the API. It is a reason to be honest about how the piece was made, and to keep your editing process real — actual human review, fact-checking, and rewriting where the draft needs it. Google’s guidance on AI content has consistently been about quality and usefulness, not about banning AI assistance outright; see our summary of Google’s AI content fact-check guidance.

For product copy, remember that short text may carry no detectable mark. Do not assume a 40-word product description is watermarked just because it came from the API, and do not assume it is clean just because it is short. The x-watermark header is the only per-response fact you have — log it, and do not guess from length alone.

For client deliverables, the stakes are highest. If a client is paying you for writing and you use the Grok API to draft it, the client has a right to know — and, after November 7, the deliverable may carry a detectable mark whether or not you mention it. Tell them upfront, in the brief, the invoice notes, or a short line in the delivery email.

Instagram and Facebook have their own AI-label rules, separate from the EU AI Act. A watermark in the text does not automatically satisfy a platform’s labeling requirement, and a platform label does not replace the watermark. If you post Grok-written text to a platform that asks you to label AI-generated content, follow that platform’s rule. Do not assume the invisible watermark counts as a label — it does not, because your audience cannot see it.

A worked example: one creator pipeline, start to finish

You draft blog posts with the Grok API, edit them yourself, and publish them on your own blog. Here is what happens before and after November 7.

Step 1: Draft. You send a prompt to grok-4.6 through the Chat Completions API and get back a 1,800-word draft. Before November 7, it carries no watermark. After November 7, it does — and the response headers include x-watermark: true. You save the draft and the header value together in your content log.

Step 2: Edit. You fix two factual errors, cut a repetitive section, rewrite the introduction in your own voice, and add a personal example. The piece is now a mix of Grok-drafted text and your own writing. The watermark signal survives in the parts you kept verbatim, and is weaker or absent in the parts you rewrote. No detector can tell you, sentence by sentence, which is which — it can only give a statistical read on the passage as a whole.

Step 3: Publish. You publish the edited piece. Nothing visible changes: no badge, no footnote, no change to how the page renders. A reader sees a normal article; a detector with access would likely read it as watermarked, because enough of the original draft survives.

Step 4: What a platform or client can and cannot detect. A platform cannot detect the watermark unless it has detector access from SpaceXSI — and, as of November 6, that access is by application only. A client who asks whether AI was used will not get the answer from the watermark. They will get it from you, if you tell them. The watermark is a backstop for detection, not a substitute for disclosure.

That pipeline is the normal case, and it is not a problem — as long as you are honest about step 1 and keep a record of it. The problem only starts when step 1 is hidden and someone else discovers it later.

The disclosure checklist for client work

If you do client work with Grok-drafted text, use this checklist on every deliverable.

• [ ] Name the tool and the model. Write down that the draft used the Grok API, and which model — for example, grok-4.6 — produced it. “AI-assisted” is vague. The model name is specific, and specificity builds trust.
• [ ] Say what the AI did and what you did. One line is enough: “Grok drafted the first version; I fact-checked, rewrote the introduction, and added the client examples.”
• [ ] Log the x-watermark header. For each API response that feeds a deliverable, record the header value alongside the response — true means watermarked, anything else or no header means not watermarked. Keep that log with the project file.
• [ ] Note the date. After November 7, the default is watermarked. Before that date, it was not. A one-line date note prevents confusion later.
• [ ] Check the client’s own rules. Some clients ban AI drafting outright. Some require disclosure. Some do not care. Ask before you start, not after you deliver.
• [ ] Check the platform rules where the text will live. If the deliverable is going on Instagram, Facebook, or another platform with AI-label rules, confirm who applies the label, and apply it.
• [ ] Keep the opt-out decision separate. Do not opt out for a client project unless a team admin has made that call, confirmed a legal exception applies, and accepted sole responsibility — in writing, before the work starts. Never opt out quietly to make a deliverable look human-written.
• [ ] Keep the record. Save the checklist, the header log, and the client’s disclosure acknowledgment with the project. If a question comes up months later, you want the answer in the file, not in your memory.

That checklist is not legal advice. It is a working habit. If you have a lawyer or a compliance contact, show it to them and adjust it to your contracts.

What to do this week

You do not need to rebuild your workflow. Make four small changes before November 7.

1. Log the x-watermark header. If you already log API responses, add one field: the value of the x-watermark response header. Record it alongside each logged response, with the model name, the date, and the project it belongs to. That gives you an audit trail: for any piece of text, you can say when it was generated, by which model, and whether it was watermarked. If you do not log responses today, start with a simple file or spreadsheet.

2. Decide per model, in writing. For each of the seven models you actually use, write down a one-line decision: watermarked by default, or opted out — and if opted out, why, who approved it, and what legal exception was confirmed. Writing it down forces the conversation the opt-out process is designed to require.

3. Document the decision where the team can find it. If you work alone, put it in your API notes. If you work on a team, put it where the other admins can see it, and make sure everyone who calls the API knows the default changed on November 7.

4. Tell your clients before they ask. If you have active client work that uses Grok drafts, send a short note this week: the API now watermarks text by default, your process already includes human editing and disclosure, and nothing about the deliverable changes except the audit trail.

For background on Grok itself, see our guide to Grok Voice 3.0 for small business voice agents and our breakdown of xAI, X, and Grok subscription plans.

Two questions are still unanswered as of November 6: whether the Grok apps, Grok Imagine, or the Vertex AI and Foundry platforms will be covered, and who can get detector access. For an independent overview of the change, see this write-up. Build the logging habit now, and update your notes as those answers arrive.

Frequently Asked Questions

Does the Grok watermark change how my text looks?

No. The watermark is statistical — it slightly shapes which words the model picks. There are no visible characters, no hidden text, and no change a reader can spot. Longer text carries a stronger signal; very short text may carry no detectable mark at all.

How do I know if a specific Grok API response was watermarked?

Check the HTTP response header x-watermark. A value of true means that response was watermarked. Any other value, or no header at all, means it was not.

Does watermarking apply to the Grok app or Grok on X?

Not as of the November 7 announcement. This change is API only. SpaceXSI has announced no date for watermarking text in the Grok app, Grok on X, or the Grok website.

Can I opt out of Grok text watermarking?

Only a team admin can, per model, in the SpaceXSI Console — and only after confirming a legal exception applies and accepting sole responsibility for marking and disclosure. The opt-out applies to every API key on the team and takes effect within about five minutes. Treat it as a legal decision, not a settings toggle, and get the decision in writing first.

Does the watermark prove a piece of text was written by Grok?

Not by itself, and not conclusively. Detection is statistical, with three outcomes: watermarked, not watermarked, or uncertain. Heavily edited, translated, or very short text may carry no detectable mark. Anthropic gives the same warning about Claude’s marking — a mark “is not fully conclusive.”

Does this affect images, video, or voice from Grok?

No. This change covers text only. It does not apply to images or video from Grok Imagine, and it does not apply to voice. Other marking rules may apply to those media types separately, so check the current documentation for each surface you use.