Findivoza

November 9, 2026

Findivoza

Grok logo
Technology

Grok Now Watermarks Its AI Text by Default: What Changed on 7 November and Why It Matters

Faizan Altaf

From 7 November 2026, every piece of text that Grok writes through its developer API carries a hidden mark. You cannot see it. It adds no words, no symbols and no hidden characters. But the mark is there, woven into the very words the model chooses — and with the right tool, it can be detected later.

The change comes from xAI, the company behind Grok, which now appears under the name SpaceXSI in its developer documentation. It applies to seven Grok models at once, it is switched on by default, and the reason given is simple: a new European law says AI-generated content must be marked so machines can recognise it.

This article explains what changed, which models are affected, how the invisible watermark works, whether you can switch it off, and what it means for developers, businesses, students and ordinary readers.

What exactly changed on 7 November?

On 7 November 2026, xAI switched on text watermarking by default for its Grok API. From that day, text generated by seven Grok models through the API includes an invisible watermark. Developers did not need to change anything: there is no new setting to send, no new parameter, and existing apps and integrations keep working exactly as before.

The watermark applies on every xAI API endpoint, including the global endpoint at api.x.ai and the company’s regional endpoints. It covers both of the company’s text generation APIs — the Responses API and the Chat Completions API — and it covers the generated text in the response: ordinary text replies, structured outputs, and even the arguments in tool calls.

One important limit: the change covers the developer API only. xAI has announced no date for watermarking in the Grok app, on the Grok website, or in Grok on X. So if you chat with Grok on X or in the app, nothing has changed for now.

Which Grok models are affected?

The watermark applies to seven models, from the newest flagship down through the Grok 4.20 family:

  • 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

Grok 4.7 is the current flagship model. The Grok 4.20 variants are specialist versions of the same family: one built for step-by-step reasoning, one for ordinary replies, and one that coordinates several AI agents working together.

If your product calls any of these seven models through the API and you have not touched the watermark setting, your text responses have carried the invisible marker since 7 November — without any code change on your part.

How does an invisible watermark work?

A text watermark is not a stamp or a label you can see. It is a statistical pattern hidden inside the words themselves.

When a language model writes, it does not pick the next word with certainty. At each step it weighs many possible words and chooses one. A watermarking system gently shapes those choices — favouring some words over others in a pattern that follows a secret key. The text still reads normally. The meaning does not change. But across a passage of enough length, the pattern becomes statistically detectable: a detector can look at the word choices and say whether the text was probably written by that model.

xAI’s own documentation describes it this way: 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 to the response.” The company also says the mark does not change response quality or speed.

How can you tell if text is watermarked?

For a single API response, there is a simple check. Every response from the covered models now carries a new HTTP header called x-watermark. If the value is true, that response’s output was generated with the watermark. Any other value — or no header at all — means it was not.

For text found elsewhere — say, an article or a homework essay that might have been written by Grok — the check is harder. xAI runs a detection tool that scores a passage and reports one of three results: watermarked, not watermarked, or uncertain. But access to that detector is granted by application only. It is not open to the public, and an ordinary reader has no way to check a piece of text on their own.

Can you switch the watermark off?

Yes — but not freely, and not quietly.

Only a team administrator can turn watermarking off, and only in the SpaceXSI Console, in the Team settings page. The setting is controlled per model, for the whole team, and it applies to every API key on that team. It takes effect within a few minutes.

To opt out, the admin must confirm that a legal exception under applicable law covers their use — and that they are solely responsible for any marking, disclosure or other legal requirement. In other words, switching the mark off means the team takes the legal responsibility onto itself.

There is no per-request switch. You cannot ask for a watermark on one reply and not on the next, and no individual developer on a team can override the team’s setting.

Why now? The EU AI Act deadline

The reason xAI gives is one line: “SpaceXSI watermarks text model output to comply with the EU AI Act.”

Article 50 of the EU AI Act is the transparency rule. Since 2 August 2026, it has required providers of generative AI systems to mark their synthetic content — text, images, audio and video — in a machine-readable format, so it can be detected as artificially generated. Systems already on the market before 2 August were given a grace period, which ends on 2 December 2026. Any AI system that was on the market before August must meet the marking duty by then, or face fines of up to €15 million or 3% of worldwide annual turnover — whichever is higher.

Anthropic moved first, watermarking text from three older Claude models from 3 November, with four more to follow on 16 November. xAI followed four days later, on 7 November. OpenAI launched its own text watermark, called textGrain, in October — but it is switched on by default only for ChatGPT and Codex users in the EU.

The law is European, but the effect is global. xAI has switched the watermark on at every API endpoint, not just in Europe. Anthropic has done the same worldwide. For a developer in Pakistan building an app on the Grok API, the mark applies just as it does for a developer in Paris.

Grok is not alone: the watermark race

Grok’s move is part of a wider shift that arrived almost all at once:

  • Anthropic’s Claude: Newer Claude models have carried a watermark since they launched after 2 August. Three older models — Mythos 5, Opus 4.7 and Sonnet 4.6 — started carrying it from 3 November, and four more (Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5) follow on 16 November. Opus 4.6 was delayed from the first wave to the second. Unlike Grok, Claude’s customers cannot switch the mark off — but on some older models the mark only appears when developers leave three sampling settings at their defaults, which is a loophole we reported earlier.
  • OpenAI’s ChatGPT: Its textGrain watermark is on by default for users in the EU, and optional for API customers.
  • Google’s Gemini: Google’s SynthID method, which Anthropic’s watermark is based on, has been marking AI-generated images, audio and video for longer.

The result is that by December, text from most of the big AI systems will carry some kind of hidden mark.

What this means for you

The change affects different people in different ways.

If you build apps with the Grok API

This is the group most directly affected. If your app, website or tool calls one of the seven models, its output now carries the mark by default. That is probably good news: it shows your product is playing by the rules. But check your setup — if your team had already opted out before 7 November, the setting stays as it is, and switching it off now means your team accepts sole legal responsibility for marking.

If you run a business that publishes AI-written content

The watermark does not change what your readers see. But it does change what can be proven about your content later. A regulator, a platform or a fact-checker with detection access could one day test a passage and find the Grok mark in it. If your business relies on passing AI text off as human-written, that is now a riskier game — and in the EU, it could be an unlawful one.

If you are a student or a teacher

Watermarks are often described as a tool against cheating. In practice, they are a partial one: heavy editing or paraphrasing can weaken or remove the signal, and short answers may not carry enough of it to judge. Teachers should not treat a watermark test as proof on its own. The honest rule has not changed: say when you used AI, and do your own thinking.

If you just chat with Grok

Nothing changes for you right now. The Grok app, the website and Grok on X are not covered, and xAI has announced no date for them. Your chats are unaffected, and you cannot check whether a piece of text you read online was written by Grok — the detection tool is not open to the public.

If you are in Pakistan

The change reaches you through the API, not through any local law. Pakistan has no watermarking rule of its own. But if you build products for customers in Europe, or publish content that reaches European readers, the EU rules apply to that content.

The limits of watermarks

Watermarks are a real step forward, but they are not a complete answer. Four limits are worth knowing.

First, detection is not public. A system meant to help the public spot AI text currently lets only approved organisations check for the mark. Until that changes, the transparency benefit is limited.

Second, editing weakens the signal. Paraphrasing, translating, or heavily rewriting a passage can wash out the statistical pattern. A watermark is evidence, not proof — and a missing watermark does not prove a human wrote the text.

Third, the rules differ between companies. Grok lets team admins opt out. Claude cannot be switched off, but on some older models the mark vanishes if a developer changes one sampling setting. OpenAI’s watermark is EU-only by default. A patchwork of different rules means the same text could be marked by one system and unmarked by another.

Fourth, a watermark says who wrote the text, not whether it is true. A watermarked passage can still contain errors, invented facts or biased claims. The mark tells you the author was a machine. It says nothing about accuracy, so readers still need to check facts for themselves.

What happens next

Three things are worth watching.

  1. The 2 December deadline. The EU grace period ends on 2 December 2026. Expect more providers to announce their marking plans before then, and watch for the first enforcement action against a company that misses the deadline.
  2. The Claude wave on 16 November. Anthropic’s second wave of watermarking reaches its most widely used models — Opus 4.5, Sonnet 4.5 and Haiku 4.5 — on 16 November. That will put a mark on a large share of the AI text produced through APIs.
  3. Whether detection opens up. The biggest open question is who gets to check. If xAI and Anthropic open their detectors to journalists, teachers and the public, watermarks become a genuinely useful transparency tool. If detection stays locked away, the marks will mostly serve regulators and the companies themselves. Neither company has given a date for wider access. Also watch whether xAI extends watermarking to the Grok app and to Grok on X — for now the consumer products are untouched.

Frequently asked questions

What is a Grok text watermark?

It is an invisible, statistical pattern hidden in the words Grok chooses when it writes text through its API. It adds no visible characters and does not change the meaning, but a detector can test a passage and tell whether the pattern is there.

When did Grok start watermarking text?

On 7 November 2026. From that date, text from seven Grok models — grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants — carries the mark by default when generated through the API.

Does the watermark change what Grok writes?

No, according to xAI. The company says the mark does not change response quality or speed, and adds no characters, markup or hidden text. It only shapes which words the model picks as it writes.

Can I check whether a piece of text was written by Grok?

Not as an ordinary reader. Developers can check a single API response by reading the x-watermark header — true means it was watermarked. For any other text, xAI’s detection tool is available by application only, to approved organisations such as regulators, researchers and media.

Can the watermark be turned off?

Only by a team administrator, per model, in the SpaceXSI Console — and only after confirming that a legal exception applies and accepting sole responsibility for marking. Individual developers cannot override their team’s setting.

Why is xAI doing this?

To comply with the EU AI Act. Article 50 of the Act requires AI providers to mark synthetic content in a machine-readable way. The grace period for systems already on the market ends on 2 December 2026, with fines of up to €15 million or 3% of worldwide turnover for breaches.

Does this affect the Grok app or Grok on X?

No, not yet. The change covers the developer API only. xAI has announced no date for watermarking in the consumer app, on the website, or on X.

Conclusion

On 7 November, Grok joined the quiet revolution in how AI text is marked. Seven models, one invisible pattern, switched on by default — and a clear reason: a European law that says machines must own up to what they write.

The change is invisible to readers and painless for developers. But it is a real shift: a large share of the AI text produced every day now carries a signal that can, in principle, be traced back to its source. The open questions are who gets to check for that signal, how easily it can be washed away by editing, and whether the consumer apps follow.

For now, the practical advice is simple. If you build with the Grok API, check your team’s watermark setting and know what it means. If you publish AI-written content, assume it can one day be identified as such. And if you read something online, remember that a watermark can tell you a machine wrote it — but only your own judgement can tell you whether it is true.

Sources

This article is for general information and education only. It is not legal or financial advice.

Faizan Altaf

Faizan Altaf is the founder and author of findivoza.com, a news blog covering world, Pakistan, business, technology, health and sport in plain English. He writes for general readers who want clear, trustworthy news without jargon. His work focuses on explaining what happened, why it matters, and what to watch next.

Leave a Reply

Your email address will not be published. Required fields are marked *