Grok API Watermark Explained: Invisible Mark on 7 Models From November 7, and How to Check It

Image: xAI — Grok logo (official logo, via Wikimedia Commons, public domain)

TL;DR: From 7 November 2026, text from seven Grok models carries an invisible watermark by default when generated through xAI's API. You cannot see it, your code does not need to change, and the only visible trace is an x-watermark: true response header. It is xAI's answer to the EU AI Act's transparency rule — and team admins can switch it off, model by model.

📌 KEY HIGHLIGHTS

Watermark on by default from 7 November 2026 for grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants, on every API endpoint.
Invisible by design: it shapes the model's word choices, adds no characters, markup or hidden text, and xAI says it changes neither quality nor speed.
Check any response with the x-watermark header; detection itself is statistical and available only by application.
Team admins can opt out per model in the SpaceXSI Console — but must confirm a legal exception applies and accept responsibility for marking.
Consumer Grok (the app, grok.com and X) is not covered — the change is API-only for now.
The driver is the EU AI Act: transparency duties in force since 2 August 2026, with a 2 December 2026 deadline for older systems.

What changed on November 7?

xAI — the company behind Grok, whose developer documentation now carries the name SpaceXSI — switched on invisible text watermarking for its API. From 7 November 2026, every text response generated by seven Grok models through the API carries the mark by default. The change was documented on 31 October 2026 and took effect a week later, with no announcement fanfare: developers discovered it in the docs, and the only thing that changes in your integration is a new response header.

Nothing else moves. There is no request parameter to send, no new field in the response body, and no change to pricing. Your existing code keeps working exactly as before — the watermark is applied while the model generates text, by slightly shaping which tokens it picks. That is the whole point: the text reads exactly the same, but xAI can later recognise it as machine-generated.

The seven Grok models now watermarked

The watermark covers every text-generation API — the Responses API and the Chat Completions API — on every endpoint, including the global api.x.ai address and regional endpoints. It covers ordinary text responses, structured outputs and tool-call arguments, so even the JSON your agent produces is marked.

Model	Watermarked from	How to check
grok-4.7	7 November 2026	x-watermark header
grok-4.6	7 November 2026	x-watermark header
grok-4.5	7 November 2026	x-watermark header
grok-4.3	7 November 2026	x-watermark header
grok-4.20-reasoning	7 November 2026	x-watermark header
grok-4.20-non-reasoning	7 November 2026	x-watermark header
grok-4.20-multi-agent	7 November 2026	x-watermark header

Note what is not in the list: the Grok app, the grok.com website and Grok inside X. The documentation covers the developer API only, and no date has been announced for the consumer surfaces. If you chat with Grok on X, this change does not touch you — yet.

How the invisible watermark works

Image: Wikimedia Commons — Flag of Europe (public domain). The EU AI Act is the reason behind the watermark.

A text watermark is not a stamp on the page. It is a statistical pattern hidden in the model's word choices: while generating, the model is nudged toward certain tokens over equally good alternatives, in a pattern only the watermark holder can recognise. No hidden characters are inserted, no markup is added, and nothing is visible to the reader.

Detection works the same way. xAI scores a piece of text against the watermark and reports one of three results: watermarked, not watermarked, or uncertain. Longer text gives a stronger signal — a one-line reply may come back uncertain, while a full article is much easier to call. And detection is not something you can try yourself: access to the detection API is granted by application, for regulators, researchers, media and other approved organisations. There is no public checker.

One claim deserves a caveat. xAI says watermarking changes neither the quality nor the speed of responses. That is the vendor's own statement — it has not been independently measured, and independent evaluations of similar watermarks have found detection rates far below what providers advertise. Treat the quality claim as xAI's word, not a tested fact.

How to check a response — and how to opt out

Checking is a one-liner. Read the x-watermark header on the Responses API or Chat Completions API reply: a value of true means that request's output was generated with the watermark. Any other value, or no header at all, means it was not.

Opting out is deliberately a team-level decision, not a per-request one. Only team admins can do it, in the SpaceXSI Console under Team settings, in the "Model output watermarking" row. They can switch the mark off for some or all models, and changes apply to new requests within five minutes. Every opt-in and opt-out is recorded in the team's audit log.

There are two catches. First, opting out asks the admin to confirm that an exception under applicable law covers the use, and that the team alone is responsible for any marking, disclosure or other legal requirement. Second, some models may be locked on for a team — locked models appear in the dialog without a checkbox and cannot be turned off from the console at all.

Grok vs Claude: two opposite ways to watermark

Grok is not the first to do this. Anthropic extended its own invisible text watermark to four older Claude models — Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 — from 3 November 2026, just four days before Grok's switch-on. Every Claude model released since 2 August 2026 already shipped with the mark. The two companies have chosen opposite postures, and the difference matters if you build on either.

	Grok API (xAI)	Claude (Anthropic)
Switched on	7 November 2026	3 November 2026 (older models); since 2 August for new models
Opt-out	Yes — team admin, per model, in the console	No — customers cannot turn it off
Visible signal	x-watermark: true response header	None to the user
Coverage	API only (for now)	API, apps, Claude Code, Cowork and cloud partners
Detection	By application	Private preview for eligible organisations
Known loophole	Some models may be locked on	On Opus 4.6 and Sonnet 4.6, the mark only applies at default sampling settings

OpenAI has its own method too: textGrain, announced on 5 October 2026, intended for ChatGPT and Codex in the EU. The pattern is now clear — every major lab is marking its text, each in its own way, and Grok's admin-controlled opt-out is the most flexible of the three.

Why now: the EU AI Act's December deadline

The reason is one line in xAI's docs: the company watermarks text "to comply with the EU AI Act". Article 50 of the Act requires providers of generative AI systems to mark synthetic content in a machine-readable format so it can be detected as AI-generated. Those transparency duties have applied since 2 August 2026. Systems already on the market before that date got a grace period — until 2 December 2026 — and 7 November falls 25 days before that deadline.

That timing explains the rush across the industry in the first week of November: Anthropic retrofitting older Claude models on 3 November, xAI switching Grok on by default on 7 November. The law is European, but the implementation is global — both companies apply the mark worldwide, not only to European users, because splitting outputs by region would be harder than marking everything.

What it means for Indian developers

There is no India-specific carve-out: the watermark is on for every endpoint, including the ones Indian developers call. If your product uses the Grok API — a support chatbot, a content tool, an agent that drafts emails — its output now carries the mark, wherever your users are.

Three things to do. First, check the x-watermark header in your integration so you know which responses are marked. Second, if you resell or republish Grok output — especially to customers in Europe — remember the mark travels with the text, and your own disclosure duties may follow from it. Third, if your use case genuinely needs unmarked output, decide at team level whether to opt out in the console, document the legal basis you are relying on, and note that the decision is audit-logged. India has no equivalent blanket watermarking rule in force, so the practical effect here is compliance with a European rule that arrives through the API — not through Indian law.

Should you care?

✅ Yes, if you build with the Grok API. The mark is on by default, it is invisible, and it changes nothing in your code — but it changes what your product's output provably is. Check the header, decide about opt-out now, and do not promise customers unmarked text without checking the console setting first. If you only chat with Grok on X or grok.com, nothing changes for you yet — but the direction of travel is obvious, and the app will likely follow.

FAQs

What is the Grok API watermark?
It is an invisible statistical mark that xAI adds to text generated by seven Grok models through its API, starting 7 November 2026. It shapes the model's word choices so xAI can later recognise the text as machine-generated. It adds no characters, markup or hidden text, and readers cannot see it.

Which Grok models are watermarked?
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. The mark applies to the Responses API and the Chat Completions API on every endpoint, including structured outputs and tool-call arguments.

How do I check if Grok text is watermarked?
Read the x-watermark header on the API response: a value of true means that request's output was generated with the watermark. Full detection — scoring a piece of text as watermarked, not watermarked or uncertain — is available only through xAI's detection API, which is granted by application.

Can I turn off the Grok watermark?
Team admins can switch it off per model in the SpaceXSI Console, under Team settings, in the "Model output watermarking" row. Changes apply within five minutes and are audit-logged. You must confirm a legal exception applies and accept sole responsibility for marking. Some models may be locked on and cannot be switched off.

Takeaway: Grok's text is now invisibly signed at the source. The change is small, silent and global — and for anyone building on the API, it is worth a five-minute check of the console before a customer asks about it first.

If you build on the Grok API, open your team's watermark setting in the SpaceXSI Console today and read the official Output Watermarking docs — it takes five minutes and settles whether your product's output is marked.

TechByCountry — Daily Tech News, Country by Country

Search This Blog