SpaceX SI Will Watermark Text From 7 Grok API Models Starting Nov. 7. It's On by Default, It's Invisible, and Opting Out Makes the Legal Duty Yours.

Honest fence (read first): This comes from SpaceX SI's own developer docs, read in full. It's a scheduled change for the Grok API, not something that's already live, and the docs don't say anything about the Grok apps or Grok Bot. SpaceX SI hasn't posted about it on X.

Editorial Thesis

If you build on the Grok API, the text you get back is about to carry a mark you can't see.

A new Output Watermarking page in SpaceX SI's developer docs says that starting Nov. 7, 2026, text from seven Grok API models includes an invisible watermark. It's switched on by default for every team. SpaceX SI gives one reason: "to comply with the EU AI Act."

That's the claim, and here the proof is unusually solid, because the company wrote it down in its own docs. What isn't proven is how well the watermark holds up once people edit the text, and who will actually be able to check for it.

Honest Verdict

• Confirmed (SpaceX SI docs): the Nov. 7 start, the seven models, on by default, every endpoint, the x-watermark header, the opt-out rules.
• Reported: the customer email that first surfaced the change, shared on X by a developer on Oct. 31 (via Tech City Authority).
• Not known: how robust the mark is against paraphrasing, how accurate detection is, who gets detection access, and whether the Grok apps or Grok Bot will be marked too.

What changes on Nov. 7

The docs list the seven models that will carry the watermark:

• grok-4.7, grok-4.6, grok-4.5 and grok-4.3
• grok-4.20-reasoning, grok-4.20-non-reasoning and grok-4.20-multi-agent

It covers every text generation API, including the Responses API and the Chat Completions API. That means ordinary replies, structured outputs and the arguments a model writes when it calls a tool.

It isn't limited to Europe. The docs say the watermark "is enabled on every SpaceXSI API endpoint, including the global endpoint, https://api.x.ai, and any regional endpoint."

How it works

You won't see anything in the text. The docs say the watermark is applied while the model writes: "it slightly shapes which tokens the model picks, so it adds no characters, markup, or hidden text to the response."

In plain terms, the model leans toward a hidden pattern of word choices. Software that knows the pattern can score a passage later.

• Detection is statistical. SpaceX SI scores text as "watermarked, not watermarked, or uncertain." Longer text gives a stronger signal.
• Detection isn't public. Access to the detection API "is granted by application" through a form.
• No code changes. There's no request parameter. A response header, x-watermark: true, tells you that a reply was marked. "Any other value, or no header at all, means it was not."
• Quality claim. The docs say watermarking "doesn't change response quality or speed." That's the company's statement; we haven't seen independent tests.

The opt-out comes with strings

Team admins can turn watermarking off in the SpaceX SI Console, model by model. The setting covers every API key on the team, takes effect within 5 minutes, and every change lands in the team's audit log.

But the switch isn't free. Opting out asks the admin to confirm "that an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure, or other legal requirement."

And some teams won't get the choice at all: "Some models may be locked on for your team." Locked models show in the dialog without a checkbox. The docs don't say which teams or models get locked, or which uses count as an exception.

Why this matters

• For developers: anything you ship that's written by Grok through the API (emails, product copy, support replies) can later be flagged as machine-written by anyone with detection access. If your customers expect that text to read as human-written, that's a disclosure conversation to have now, not after Nov. 7.
• For the industry: SpaceX SI isn't alone. The Tech Signal notes Anthropic has signed the EU's voluntary transparency code and watermarks Claude models released in the EU since Aug. 2. We covered Claude's version here: a watermark is a signal, not proof of who wrote something.
• For readers: a "not watermarked" result won't prove a person wrote something. Opted-out teams, other AI models and edited text all fall outside it.

Claim vs proof

What to watch next

• Whether x-watermark: true shows up on real responses from Nov. 7.
• Any SpaceX SI post or changelog entry explaining exceptions or locked models.
• Who gets detection access, and any independent tests of accuracy after editing.
• Whether the same marking reaches the Grok apps or Grok Bot.

The video

ArcCraft Media's short explainer (Nov. 1, 4:29) walks through the Nov. 7 change and cites the same docs page. It's a small channel with AI narration; the facts above come from the docs. Alternate: Will Francis, "AI Watermarks Are Here to And You Can't See Them" (Nov. 2, 6:11), which explains text watermarks using Claude's version.

Bottom line

From Nov. 7, text from seven Grok API models carries an invisible watermark by default, and SpaceX SI says it's doing it for the EU AI Act. Developers don't have to change code, but opting out makes any legal labeling duty theirs. It's a real, documented change. How well the mark survives editing is still unproven.