I build things with AI APIs, so every Monday I skim the developer docs of the tools I use. On Saturday, November 7, one quiet doc update turned out to be anything but quiet: xAI — the company behind Grok, which now appears as SpaceXSI in its own developer documentation — flipped a switch, and from that day on, text generated by seven Grok models carries an invisible watermark by default.
No announcement banner. No new request parameter. No change to the response body. But the words coming back from the API are now marked as machine-made — and if you build products on top of Grok, that matters before December 2, when the EU's transparency deadline lands.
What actually changed on November 7
xAI's developer page "Output Watermarking" (updated October 31, 2026) says text generation from these seven models carries an invisible watermark by default from November 7, 2026: 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.
That is the flagship line plus the whole 4.20 family. The marking applies across both text APIs, the Responses API and the Chat Completions API, on every SpaceXSI endpoint, including the global api.x.ai host and regional endpoints — structured outputs and tool-call arguments included.
You do not have to change a single line of code: no new parameter, no new field in the JSON. The watermark lives in the token selection itself — the statistical pattern of which words the model picks — so the text reads exactly the same to a human. The only visible trace is a new response header: x-watermark: true in the raw HTTP response means that request's output was watermarked. Any other value, or a missing header, means it was not.
Most of us never see raw headers — our SDKs parse the JSON body and hand us the text. So if you want to know, per request, whether your output was marked, you have to start capturing headers in your API client.
One scope note: this is an API change only. The docs say nothing about Grok on X, the grok.com chat UI, or the consumer app, and nothing about Grok Imagine images or video, which already carry a separate watermark.
How an invisible text watermark actually works
A language model does not pick words the way you or I do. At every step it produces a probability distribution over the next token, then samples from it. A statistical watermark biases that sampling very slightly — nudging the model toward certain tokens over others according to a secret pattern — invisible in any single sentence, but detectable across a longer passage if you know the pattern.
That is the key trade-off. A watermark is not a stamp on the page: you cannot open a Grok reply and see the mark, and no free online detector can reliably spot one. For now, only xAI — the party holding the secret pattern — can say with confidence whether a given piece of text carries it. The x-watermark header is your per-request receipt; the statistical test is theirs.
That is the nature of the technique: the watermark makes text recognizable as model-generated by the provider, not provably AI-written to any random reader. If you are building a product that promises users it can detect AI text, this watermark does not give you that product. One caveat: xAI states watermarking does not change response quality or speed — the vendor's claim, with no independent measurements published yet.
Why now: the EU AI Act and the December 2 deadline
xAI gives a one-line reason: it watermarks text "to comply with the EU AI Act."
Article 50 of the EU AI Act requires providers of AI systems that generate synthetic text, audio, images, or video to mark their outputs in a machine-readable format detectable as artificially generated. Those transparency obligations have applied since August 2, 2026. For systems already on the market before then, the transition window runs until December 2, 2026 — November 7 falls 25 days before that deadline.
The change is global rather than Europe-only: it is cheaper to ship one behavior everywhere than to maintain a regional fork of your inference stack. If you are a developer in Pakistan, the US, or anywhere outside the EU, do not assume this does not apply to you. Your Grok API output is watermarked today, wherever your servers sit.
Grok is not alone. Anthropic has announced a second round of text watermarking starting November 16. The difference is the default: Anthropic's move is framed as an experiment, while xAI has made watermarking the default for all covered outputs. Several outlets call it the first confirmed default-on watermark policy at a major frontier API. The big labs are moving from "we could mark our outputs" to "we do mark our outputs unless you opt out."
What this actually means for developers
If you call the Grok API in a product, your output is now marked. Your users will not notice — the text reads the same — but the provenance of what your product produces has changed. Blog posts, support replies, product descriptions now carry a machine-detectable signature. Your team should know that, not discover it in a compliance review six months from now.
If you resell AI-generated content, talk to your customers. If your agency delivers "human-written" articles or your tool sells SEO content from a pipeline calling grok-4.7, that content is watermarked at the source from November 7. I am not a lawyer, but selling watermarked AI text as human work was already dishonest, and now it is also detectable. Disclose AI involvement.
If you run a WordPress site with AI-assisted content, this is disclosure handled for free. A watermark does not replace a visible note on the page, but the content itself now carries a machine-readable mark. Pair it with a clear "written with AI assistance" line, and you are doing transparency properly instead of performing it.
If you build detection tooling, do not over-claim. Only xAI can run the statistical test, and there is no public detector. Do not promise users you can spot watermarked Grok text without a real verification path — the x-watermark header from the generation request, or an arrangement with the provider. False confidence in AI detection is worse than no detection at all.
If you heavily edit Grok output, expect the signal to degrade. Statistical watermarks survive light editing and fade under heavy rewriting — if your pipeline rewrites 80 percent of a Grok draft, do not assume the watermark survives. Keep your own logs of what was AI-generated; a watermark is a backup, not a database.
The opt-out: how it works, and why you should think twice
Yes, you can turn it off. Team admins can switch watermarking off model by model in the SpaceXSI Console, under a "Model output watermarking" control in Team settings. But the process is deliberately not a casual toggle.
Opting out asks you to confirm that a legal exception covers your use, and that you alone are responsible for any marking or disclosure. The moment you switch it off, the compliance burden moves from xAI to you. The change applies to every API key on the team, takes effect within about five minutes, and is recorded in the audit log. Some models may also be locked on.
My honest take: most teams should leave the default alone. If you have a specific, lawful reason to need unmarked output — and legal advice saying an exception covers you — the opt-out exists for you. But opting out because the toggle is there trades a small invisible signal for a legal responsibility you probably do not want. The default is the compliant path.
The bigger picture: provenance is becoming infrastructure
For two years, the argument about AI-generated text has been mostly about detection: can we build a detector that catches it? The detectors kept failing — paraphrasing broke them, false positives embarrassed them. Watermarking flips the problem around: instead of asking the reader to detect AI text after the fact, you ask the generator to mark it at the source. That is why the default matters so much. An opt-in watermark is a feature almost nobody enables; a default-on watermark is infrastructure. Once one major lab makes it the default, the pressure on the others becomes hard to resist. For those of us building on WordPress and Elementor, that matters: our ecosystem runs on content, and the line between human-written and AI-assisted is exactly where trust lives.
What I would do this week
If you use the Grok API, here is the checklist I am running:
- Check which models you call. If any are on the seven-model list, your output is watermarked now. Confirm it, don't assume it.
- Start capturing the
x-watermarkheader. Log it alongside each API call, the same way you log latency and token counts. - Tell your team. A one-line message in your team channel is enough — the worst outcome is someone discovering this during a client call.
- Review your content disclosure. If your product publishes AI-generated text, make sure the human-facing disclosure matches what the watermark already says.
- Decide on the opt-out deliberately, or not at all. If nobody on your team has a documented legal reason to opt out, leave the default on.
- Watch your quality metrics. xAI says watermarking changes neither quality nor speed. Believe it, but verify it on your own traffic.
None of that is dramatic — a header logged, a disclosure line added, a default left alone.
Conclusion
On November 7, 2026, Grok's API started marking its words — not with a visible badge or a new field in your JSON, but with a statistical pattern woven into the text itself and a small header on the response. Seven models, both text APIs, every endpoint, no code change required, and a December 2 regulatory deadline as the reason.
This is one of those changes that will look obvious in hindsight. The EU AI Act said AI content must be marked as machine-made. xAI went first with a default-on answer, Anthropic follows on November 16, and the rest of the industry must now decide whether to match the default or explain why it didn't. For developers, the response is a short checklist: know which models you call, log the header, tell your team, keep your disclosures honest, and leave the default alone unless you have a real reason to change it. Do those things and this story costs you an afternoon.
The web is about to get much better at knowing which words a machine wrote. That is a good thing — as long as the people building on it are honest about it too.
Frequently Asked Questions
Which Grok models are watermarked by default? Seven models from November 7, 2026: 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. It covers the Responses API and the Chat Completions API on every SpaceXSI endpoint, including structured outputs and tool-call arguments.
How do I know if a Grok API response was watermarked? Check the x-watermark response header on the raw API reply. A value of true means that request's output was generated with the watermark; any other value, or a missing header, means it was not. The response body looks the same, so capture headers in your API client.
Can I turn the Grok watermark off? Team admins can opt out per model in the SpaceXSI Console, under "Model output watermarking" in Team settings. Opting out requires confirming that a legal exception covers your use and that you alone are responsible for any marking or disclosure. The change applies to every API key on the team, takes effect within about five minutes, and is recorded in the audit log. Some models may be locked on.
Does watermarking change the quality or speed of Grok's replies? xAI states that it does not. That claim comes from the vendor's own FAQ and has not been independently measured yet, so watch your own latency and quality metrics if you run Grok at volume.
