According to xAI’s developer documentation, Grok API watermarking switched on by default on 7 November 2026. From that date, text generated through xAI’s API by seven Grok models carries a statistical watermark, and developers need to change nothing to receive it. The change is set out in xAI’s developer documentation, updated 31 October, which says the watermark is “on by default” and can be switched off per model.
The move lands weeks ahead of a European deadline. For anyone building on the Grok API — or reading text that may have come from it — the practical effects are small today, but the direction is clear: machine-readable marks on AI text are becoming standard practice.
What changed on 7 November
According to xAI’s documentation, the watermark now applies to outputs from every text generation API: the Responses API and the Chat Completions API. It covers ordinary text responses, structured outputs and tool-call arguments, and it is enabled on every endpoint, including the global endpoint at api.x.ai and any regional endpoint.
The watermark itself is statistical, not visible. It adds no characters, markup or hidden text to the response, so nothing in the text a developer receives looks any different. xAI states that watermarking does not change response quality or speed — that is the company’s own claim, and no independent testing of it has been published.
One detail worth noting: the documentation now brands the company as SpaceXSI, formerly xAI, while the developer domain remains x.ai. The switch-on date was announced in advance — the docs page was updated on 31 October with the 7 November date — so this was a planned change, not a quiet one.
Which models are covered
The change covers exactly seven models:
- 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
No other Grok model is named in the documentation. If a model you use is not on that list, the docs give no indication that its output is watermarked under this change.
How to check whether a response was watermarked
Developers can check each response directly. Read the x-watermark header on the Responses API or Chat Completions API response: a value of 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.
No code change is needed to receive watermarked text — the documentation states there is no parameter to send, and an integration works the same way with or without it. That makes the header check the simplest way to confirm the change is live for a given model and endpoint, and it is the check to run before telling anyone downstream that your output is marked.
Detection of the watermark in a piece of text is a separate matter. It is statistical, with three possible outcomes — watermarked, not watermarked or uncertain — and longer text gives a stronger signal. Access to xAI’s detection API is granted by application only; it is not public, and xAI has published no accuracy figures, false-positive rates or figures for how well the mark survives paraphrasing or translation.
Why xAI switched it on
The reason given in the documentation is compliance with the EU AI Act. Article 50(2) of Regulation (EU) 2024/1689 requires providers of AI systems that generate synthetic text, audio, images or video to mark the output in a machine-readable format, so it can be detected as artificially generated. Those transparency duties apply from 2 August 2026.
Systems already on the market before that date have a transition period for the marking requirement: Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, gives them until 2 December 2026. Legal analysis from Houthoff confirms the transition period. The 7 November switch-on therefore lands about 25 days ahead of the deadline.
The deadline is 2 December 2026 — some secondary commentary prints 2027, but that date applies to high-risk systems under the Act, not to the marking duty. Getting the year right matters for any team planning around it: the marking obligation is weeks away, not a year away.
Who can opt out, and how
Watermarking can be switched off, but only by team administrators and only per model. In the SpaceXSI Console, under Team settings, a Model output watermarking row offers On all, Off all or Custom, with the choice made model by model. Changes apply to new requests within five minutes.
Opting out is not a casual toggle. An admin must confirm that an exception under applicable law applies to their use, and that the team is solely responsible for any marking, disclosure or other legal requirement that follows. Some models may be locked on for a given team and cannot be switched off from the Console at all. Every opt-in and opt-out is recorded in the team’s audit log, so the decision leaves a trail.
Teams that have been weighing whether to opt out will want to settle that question with their legal or compliance colleagues before the December deadline, not after it. The same control-your-content question comes up on the training side too — our guide to opting AI training out with robots.txt walks through the other side of that decision.
The Grok app, grok.com and X are not covered
The documentation covers the API only. It says nothing about the Grok app, grok.com or Grok on X, and nothing about Grok Imagine image or video output. xAI has announced no watermarking date for those consumer surfaces, so this change should not be read as covering them. Anyone writing that consumer Grok is watermarked would be going beyond what the sources support.
Claude is moving the same way
xAI is not the only provider marking model text. Anthropic has watermarked Claude text since August 2026, using a method based on Google DeepMind’s SynthID-Text, and it extended the coverage on 3 November to three more models: Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6.
Anthropic’s explainer, published 14 August and updated 14 October, sets out the same practical shape as xAI’s: no extra tokens, no price change and no user or organisation information in the mark. Anthropic applies the mark globally, saying it has no durable way to limit it by region.
Two limits matter. First, a detected mark shows only that Claude was probably involved in producing a piece of text — it is not proof of authorship, and the absence of a mark proves nothing, since older models, heavy editing or short text can all defeat it. Second, on Claude Sonnet 4.6, text from a request that sets temperature, top_p or top_k to a non-default value might not carry the mark at all. Detection itself is restricted: Anthropic’s detection API is in private preview for eligible organisations such as regulators, law enforcement, media, fact-checkers, researchers and educational organisations.
Rollout of the November extension on Amazon Bedrock completes by 10 November, according to Anthropic’s help pages.
What to do now
For developers building on the Grok API, the immediate step is a single test call: read the x-watermark header on the response and record what it says. That one value confirms whether your integration is already covered, and it is the check to repeat for each model you use, since coverage is per model.
Next, decide your opt-out position before the 2 December deadline, and do it with legal input rather than alone — the confirmation an admin must give carries real responsibility. If you pass Grok output on to readers or customers, say plainly in your own documentation whether it is watermarked; the mark is invisible, so nobody downstream can tell without being told.
For readers, the lesson is narrower. A watermark can support a claim that text came from a model, but it cannot prove it on its own, and its absence proves nothing at all. If you want to know what AI chatbots can record about you in the first place, our AI chatbot privacy guide sets out the questions worth asking.
The open question now is whether xAI extends marking to the Grok app before the December deadline — and how quickly other providers follow with defaults of their own.
How do I check if Grok API output is watermarked?
Read the x-watermark header on the Responses API or Chat Completions API response. A value of x-watermark: true means that request’s output was generated with the watermark; any other value, or no header at all, means it was not. Check each model you use, since coverage is per model.
Does the watermark change the text I receive?
No. It adds no characters, markup or hidden text, so the response looks exactly the same as it would without it. xAI also states that watermarking does not change response quality or speed, though that is the company’s own claim and no independent testing of it has been published.
Is the Grok app on X watermarked too?
Not under this change. xAI’s documentation covers the API only and says nothing about the Grok app, grok.com or Grok on X, and no watermarking date has been announced for those surfaces. Do not assume text from the Grok app is watermarked.
Can I turn Grok API watermarking off?
Team administrators can, per model, in the SpaceXSI Console under Team settings. They must confirm that a legal exception applies to their use and accept sole responsibility for any marking or disclosure duties. Some models may be locked on, changes apply to new requests within five minutes, and every change is recorded in the team’s audit log.
Why is this happening now?
The stated reason is compliance with the EU AI Act. Article 50(2) requires providers of AI systems that generate synthetic text to mark outputs in a machine-readable format, and the transition deadline for systems already on the market is 2 December 2026. Anthropic made a parallel move on 3 November, extending Claude watermarking to three more models.
Ahesanur Asun is the Founder and Editor of UptoFeed, where he publishes updates on
AI news, artificial intelligence tools, and emerging technology. He focuses on explaining complex AI developments in a clear and practical way so readers can quickly understand and use the latest innovations.
