Grok Now Watermarks Your API Text by Default — What Developers Must Do Before 2 December

If you build on the Grok API, your output changed on 7 November 2026 — and you probably were not told. xAI switched on invisible text watermarking by default across its API, so every text response from seven Grok models now carries a statistical mark that identifies it as machine-generated. No code change was needed on your side. No email landed in your inbox. It simply went live, three weeks before the European Union’s transparency deadline.

Quick answer: From 7 November 2026, Grok API text is watermarked by default. Which models are covered, how to check the x-watermark header, and how to opt out by 2 December.

This is the first time a major frontier lab has made watermarking the default on its API rather than an opt-in. OpenAI’s equivalent stays off by default. Anthropic’s second watermarking round does not start until 16 November. If you ship Grok-powered features, the default now works for compliance — and the opt-out works against it. Here is what is marked, what the mark can and cannot prove, and the five checks to run before 2 December.

What changed on 7 November 2026

According to xAI’s own developer documentation, watermarking is now enabled by default on every SpaceXSI API endpoint, including the global endpoint and any regional endpoint. It applies to text output from text generation requests for each model that has watermarking turned on for your team. There is no request parameter to send, and the response format is unchanged apart from one new header.

The seven models covered are:

ModelWatermarked from 7 Nov 2026
grok-4.7Yes, by default
grok-4.6Yes, by default
grok-4.5Yes, by default
grok-4.3Yes, by default
grok-4.20-reasoningYes, by default
grok-4.20-non-reasoningYes, by default
grok-4.20-multi-agentYes, by default

That is effectively the full current production line-up for most teams. The watermark covers both the Responses API and the Chat Completions API — including structured outputs and tool-call arguments, not just ordinary chat text. If your pipeline feeds Grok output into another system, the mark travels with the words.

One boundary matters: this change is API-only. The consumer Grok experience on X, grok.com and the mobile apps is not covered by these docs, which describe API endpoints only. Developers are carrying the compliance weight here while end users see nothing different.

How the watermark actually works

The watermark is invisible. It is applied while the model generates text, by slightly shaping which tokens the model picks, so it adds no characters, no markup and no hidden text to the response. Nothing in the wording looks different, and nothing needs to be stripped out downstream.

The only visible technical change is a new response header. Read the x-watermark header on the Responses API or Chat Completions API response: 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. That header is your audit trail — log it in your monitoring pipeline if a compliance question ever lands on your desk.

Detection itself is statistical. SpaceXSI scores a piece of text against the watermark and reports whether it is watermarked, not watermarked, or uncertain. Longer text gives a stronger signal; short replies are harder to call. Access to the detection API is granted by application — you submit a request through a form, and there is no public detector you can run yourself today.

xAI says watermarking has no impact on response quality or speed. That is the vendor’s own claim, and it has not been independently measured. Treat it as a claim, not a test result.

What the watermark can — and cannot — prove

A watermark is a compliance measure, not a lie detector. It is worth being precise about what it does and does not establish, because the gap between the two is where teams get into trouble.

What it can do: it gives a statistical signal that a piece of text was generated by a watermarked Grok model, which is exactly what the EU AI Act’s marking obligation asks providers to supply. It survives copying and pasting, because the pattern lives in the word choices themselves rather than in hidden characters.

What it cannot do: it cannot prove who prompted the model, how much a human edited the output, or whether a passage is accurate. A missing watermark does not prove a human wrote the text — the output may simply come from an opted-out model, a short reply, or a different provider. And detection is weak in practice: reporting by byteiota, citing a 2026 independent evaluation, puts detection rates at 17 to 30 percent on pristine, unmodified watermarked text, with Google’s SynthID-Text detecting only 20 percent of its own watermarked output and falsely flagging 5.4 percent of human-written text. Paraphrasing with a second model defeats most schemes outright.

The honest read is that statistical watermarking is a reasonable compliance measure and a poor detection tool. The mark gets applied; almost nobody is reliably reading it.

That does not make the default meaningless. It makes it a legal signal, not a forensic one — and your product copy, your terms and your support scripts should describe it that way.

How xAI compares with OpenAI and Anthropic

The contrast is the story. All three labs are answering the same EU rule, but they have chosen opposite defaults.

xAI (Grok API)OpenAIAnthropic (Claude)
API defaultOn by default from 7 Nov 2026Off by default; opt-in worldwide from 5 Oct 2026On by default for covered models; no opt-out
Coverage7 Grok models, all text endpointsSelect models (API); ChatGPT and Codex in the EUModels released since 2 Aug 2026; older models retrofitted in waves
Opt-outYes, per model, team admins onlyN/A — already off unless you opt inNo — but the mark only applies at default sampling settings on older models
Detector accessBy applicationApproved researchers and expert organisationsPrivate preview
MethodStatistical token shapingtextGrain, statistical token shapingSynthID-style statistical pattern

OpenAI announced its approach on 5 October 2026: an invisible watermark called textGrain for eligible ChatGPT and Codex text in the EU, rolling out over the following weeks, while API customers anywhere in the world can switch it on for select models. OpenAI said plainly it is not making text watermarking a global default at launch.

Anthropic’s first watermarking wave was scheduled for 3 November 2026, covering Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6, with Opus 4.6 delayed to 16 November. A second wave on 16 November adds Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. The catch, reported by The Decoder from Anthropic’s customer email: on those models the watermark applies only when temperature, top_p and top_k are left at their defaults — change any one and the mark does not apply. Anthropic’s own guidance is blunt: if you rely on the watermark for compliance, keep those settings at their defaults.

xAI is therefore the first major frontier API to make the watermark the default — and the only one of the three that lets a team switch it off per model.

Why now: the EU AI Act and the 2 December deadline

The stated rationale in xAI’s documentation is compliance with the EU AI Act. Article 50’s transparency obligations have applied since 2 August 2026: providers of systems that generate synthetic text must mark their output in a machine-readable way. The European Commission said on 24 July 2026 that a limited grace period applies only to AI systems placed on the market before 2 August 2026 — and only for the marking and detection obligation under Article 50(2). Those systems must comply from 2 December 2026. That is three weeks from the day xAI flipped the switch.

Penalties for non-compliance can reach €15 million or 3 percent of global annual turnover. Every major provider has chosen a global implementation rather than watermarking EU traffic and leaving the rest untouched, so Article 50 has effectively become the de facto international baseline for AI content marking — even for teams with no EU users at all.

That global-default pattern is worth noting for anyone who assumes this does not apply outside Europe: check your provider agreements, because your provider has already decided for you.

What developers should do before 2 December

If you have Grok API integrations in production, run these five checks now, while the grace period is still open:

  1. Log the header. Start recording the x-watermark response header in your monitoring pipeline. It is the only per-request evidence you have that a given output was marked, and it costs nothing to capture.
  2. Confirm your console settings. Open console.x.ai and check the watermarking row for each model your team actually uses. The default is on, but a previous admin may already have changed it — do not assume.
  3. Test your sensitive pipelines. If your product depends on structured outputs or tool-call arguments, test output quality and tool-call accuracy now. The watermark works by nudging token choices, and subtle shifts matter most where downstream code parses the text.
  4. Decide your stance per model, in writing. Opting out is a legal assertion, not a settings tweak — document the exception you are relying on and who approved it, before the deadline rather than after a question arrives.
  5. Fix your product copy. If your terms or help pages say your AI output is “undetectable” or “indistinguishable from human writing”, that claim is now wrong for Grok API output by default. Update it before a customer or a regulator notices first.

How to opt out — and what it costs you

Team admins can turn watermarking off for some or all models in the SpaceXSI Console. The setting applies to every API key on the team, and changes take effect on new requests within five minutes. If you need different watermark policies for internal tools versus customer-facing products, xAI’s answer is to split them into separate teams in the console.

The friction is deliberate. To opt out, a team admin must confirm that a legal exception applies to your use and accept sole responsibility for any remaining marking, disclosure or other legal requirement. Every opt-in and opt-out is recorded in your team’s audit log. And some models may be locked on for your team — locked models appear in the dialog without a checkbox, and the only route is to contact support.

That design is the point: the default is compliance, and opting out means asserting you have a legal reason not to comply. Make that decision with your lawyer in the room, not in a settings panel at midnight.

Our take

xAI did the right thing in the wrong order. Making watermarking the default is the strongest compliance posture of any frontier API — stronger than OpenAI’s opt-in and cleaner than Anthropic’s sampling-settings loophole. But shipping it silently, with no email to API customers and no public announcement on the day, means the teams most affected are the last to know. A default that nobody notices is a compliance win and a trust loss at the same time.

The practical verdict for developers is simple: leave it on unless you have a documented legal reason to switch it off, log the header from today, and treat the watermark as what it is — a marking obligation satisfied, not a guarantee that anyone can detect your output later. The 2 December deadline is the moment the marking obligation becomes enforceable for older systems; the verification infrastructure that would make the mark useful — newsrooms, platforms, schools actually checking it — barely exists yet. Build for the deadline, not for the detector.

Frequently Asked Questions

Does the watermark change what Grok writes?

Not visibly. xAI says watermarking has no impact on response quality or speed, though that claim has not been independently measured. The watermark works by slightly shaping which tokens the model picks, so the text reads the same — the pattern is statistical, not visible.

Does it apply to Grok on X or grok.com?

No. This change covers the Grok API only — the Responses API and the Chat Completions API. Consumer Grok on X, grok.com and the mobile apps operates under separate policies and is not covered by the watermarking documentation.

How can I tell if a response was watermarked?

Read the x-watermark response header. A value of true means the output for that request was generated with the watermark; any other value, or no header at all, means it was not. Detection of the mark inside the text itself requires SpaceXSI’s detection API, which is granted by application.

Can I turn watermarking off?

Yes, per model, if you are a team admin — in Team Settings at console.x.ai. You must confirm a legal exception applies and accept sole responsibility for any remaining disclosure obligations. Changes take effect within five minutes and are recorded in your team’s audit log. Some models may be locked on and cannot be turned off from the console.

Does a missing watermark prove text was written by a human?

No. A missing watermark only means that particular text was not watermarked — it may come from an opted-out model, a short reply, a different provider, or a model whose sampling settings bypass the mark. OpenAI says the same about its own textGrain watermark: absence of a mark proves nothing about authorship.

Sources and Methodology

Sources: xAI / SpaceXSI developer documentation, “Output Watermarking” (updated 31 October 2026, read 9 November 2026) — docs.x.ai/developers/advanced-api-usage/watermarking; byteiota, “Grok API Text Watermarking Is Live: What Developers Must Do” (7 November 2026) — byteiota.com/grok-api-text-watermark; OpenAI, “Our approach to EU text provenance rules” (5 October 2026) — openai.com/index/eu-text-provenance; The Decoder, “Anthropic’s text watermark is easy to bypass on several older Claude models” (updated 4 November 2026) — the-decoder.com/anthropics-text-watermark-is-easy-to-bypass-on-several-older-claude-models.

Methodology: Every factual claim in this article was verified fresh on 9 November 2026 against the primary source (xAI’s own documentation, read in full) and at least one independent report. Vendor claims — xAI’s “no quality or speed impact”, OpenAI’s textGrain figures — are labelled as the vendor’s own. Detection-rate figures are reported by byteiota citing a 2026 independent evaluation, not verified by us. The 2 December 2026 date is the limited transition deadline for systems placed on the market before 2 August 2026, not a general deadline. This article was desk-researched; we did not test the Grok API hands-on and claim no hands-on experience.

Related reading: How to Opt Out of AI Training With robots.txt (2026) — the shared training opt-out token this watermarking change sits alongside. Anthropic FlashProve: The Cryptographic Receipt That Proves Which AI Really Answered You — the other end of the provenance story. How to Check Your Machine After the Tampered Claude Code Extension — the other developer-security check worth running this week.

Leave a Reply

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