
Grok AI Watermark: Can Grok Text Be Detected After November 7?
Starting November 7, 2026, text from xAI’s Grok API carries an invisible watermark by default. The company’s Output Watermarking docs (now branded SpaceXSI) say it covers seven 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. Developers do not need to change any code. Only a team admin can opt out, model by model, in the console. xAI says it is doing this “to comply with the EU AI Act.”
I have not run the Grok API with or without the watermark myself, and I have no access to xAI’s detector. Everything below comes from xAI’s documentation, the EU law text, and the coverage linked inline. I organized it as a timeline of what changes for builders, then a plain explanation of detection, then a framework for who should care.
What changes on November 7, step by step
Before November 7
Grok API output was plain text with no mark. If you pasted a Grok answer into a blog post, nothing in the words tied it back to xAI.
On November 7
According to the docs, the watermark switches on across every API endpoint, including the global https://api.x.ai endpoint and regional ones. It applies to the Responses API and the Chat Completions API, and it covers “ordinary text responses, structured outputs, and tool call arguments.” That last part means JSON your app asks Grok to write gets marked too, a point Nextfor also flagged.
Each response now includes an x-watermark header. A value of true means that request’s output was watermarked. Any other value, or no header, means it was not.
If an admin opts out
The setting lives at the team level, so it applies to every API key on the team. Changes reach new requests within 5 minutes, per xAI. Opting out asks the admin to confirm “that an exception under applicable law applies to your use” and that the team is “solely responsible for any marking, disclosure, or other legal requirement.” Every opt-in and opt-out is written to the team audit log. Some models may also be “locked on” for a team, with no checkbox at all.
What stays the same
xAI says watermarking does not change response quality or speed, and there is no new request parameter. TokenPost reports the change targets the API and not the consumer Grok app. Neural Brew said it found no official confirmation either way for the app or X, so I would treat the consumer side as unconfirmed.
How the Grok watermark actually works
The mark does not add hidden characters or markup. xAI says it is applied during generation and “slightly shapes which tokens the model picks.” A token is a chunk of a word. At most points in a sentence, several tokens would work fine, and the watermark nudges which one gets chosen in a pattern only xAI’s detector knows how to look for.
That has two consequences. Copying the text, reformatting it, or scrubbing invisible Unicode does not remove the mark, because there is nothing separate to remove. And detection is statistical. xAI scores text and returns one of three results: watermarked, not watermarked, or uncertain. The docs say “longer text gives a stronger signal.”
You cannot run that check yourself by default. Access to the detection API is “granted by application” through a form. So a teacher, editor, or client cannot paste a paragraph into a public xAI tool today and get an answer, based on what the docs describe.
Why the EU AI Act matters here
Article 50 of the EU AI Act requires providers of AI systems that generate synthetic text, audio, image or video to make outputs “marked in a machine-readable format and detectable as artificially generated.” Those transparency rules apply from August 2, 2026, according to the Act’s implementation timeline. xAI’s page names the Act directly as the reason, which explains why opt-outs come with a legal-exception confirmation instead of a simple toggle.
I covered OpenAI’s side of this question in Can People Tell I Used ChatGPT?, and the same logic now reaches Grok developers.
A framework: who the Grok watermark actually affects
I sort people into four groups based on two questions: do they call the API directly, and does their output end up in front of an EU audience or a client who checks?
| Group | Calls Grok API? | Watermark on? | What it means |
|---|---|---|---|
| App builders (chatbots, agents) | Yes | Yes by default | Log the x-watermark header for compliance records |
| Content businesses using Grok via API | Yes | Yes by default | Published text may be detectable by xAI later |
| Teams with a legal exception | Yes | Can opt out | Admin confirms exception, takes on disclosure duty |
| Consumer Grok app users | No | Unconfirmed | Docs cover the API only |
Here is a worked example, and it is a hypothetical. Picture a small content site that generates 60 articles a month through the Grok API, each about 1,200 words. That is 72,000 watermarked words a month, or 864,000 a year. Under xAI’s own rule that longer text gives a stronger signal, full-length articles are the easiest case for the detector, while a 30-word product blurb or a short JSON field is far more likely to come back “uncertain.” For that site, the risk sits almost entirely in its long posts, which are also its most valuable pages.
The practical takeaway I draw from that math is simple: if you publish long AI-drafted text, the watermark matters most to you, and heavy human rewriting is what changes the token pattern.
If you build agents, my guide What Is an AI Agent? explains why tool call arguments are now a big share of model output, and those are watermarked too.
The limitation most coverage skips
xAI has not published an accuracy rate or a false positive rate for its detector. Without those numbers, nobody outside the company can say how often “watermarked” is right, or how often human writing gets flagged. Nextfor made the same point.
Token-level watermarks also have a known weak spot. Paraphrasing or translating the text changes the token sequence the detector reads. Academic work on this style of watermark, such as Kirchenbauer et al. (2023), found that detection weakens as more of the text is rewritten. xAI’s docs make no claim that the mark survives editing or translation. There is also a gap in the header itself: x-watermark lives in the HTTP response, so it stays in your logs and does not travel with the text you paste elsewhere.
The opt-out raises its own question. A team admin can switch the mark off by attesting to a legal exception. Since xAI records it in an audit log, the attestation creates a paper trail, but the docs do not describe any up-front review of whether the exception is real.
For more on how detection claims hold up in practice, see the AdSense low-value content post, where I wrote about why thin AI text is a business risk regardless of detection, and How to Lock ChatGPT Chats for the privacy side of AI tools.
My judgment, with uncertainty
I think this is mostly a compliance move that will matter more to developers’ paperwork than to everyday readers in the short term. Detection is gated behind an application, accuracy numbers are missing, and light rewriting likely weakens the signal. I would put my confidence around 60% that, within a year, xAI or a partner opens some form of wider detection access, because the EU rules push in that direction and OpenAI and Google face the same pressure.
What could prove me wrong: xAI might keep detection private indefinitely and use it only for regulator requests, or EU guidance could set stricter technical standards that force a different method. If you publish Grok-written text at scale, I would plan as if it could be checked later.
FAQ
Can Grok text be detected?
Starting November 7, 2026, text from seven Grok API models carries a statistical watermark by default. xAI can score text as watermarked, not watermarked, or uncertain. Detection API access requires an application, so it is not a public tool.
Which Grok models are watermarked?
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, per xAI’s docs.
Can I turn off the Grok watermark?
Only a team admin can opt out, per model, in the console. The admin must confirm a legal exception applies. Some models can be locked on.
Does the watermark change Grok’s answers?
xAI says it does not affect quality or speed. It only nudges which tokens the model picks.
Is the Grok app watermarked too?
xAI’s docs only cover the API. Some outlets report consumer products are not affected, but I found no official confirmation.
Can editing remove the watermark?
Copying or reformatting will not. Heavy paraphrasing or translation likely weakens it, based on research on similar watermarks, though xAI has not published results.
Sources: xAI Output Watermarking docs, EU AI Act Article 50, EU AI Act implementation timeline, Nextfor, TokenPost, Neural Brew, Kirchenbauer et al., 2023.