Grok will secretly brand every text it writes, and you won't see it

Grok is about to sign every piece of text it writes, and nobody will see the signature. Starting November 7, all text from Grok models on the API gets an invisible watermark, switched on by default. Developers found out today from a company email, and the docs are already updated.

Seven models are covered: grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three flavors of grok-4.20 (reasoning, non-reasoning and multi-agent). Every future Grok model will ship with the mark from day one. It applies not only to plain text but also to structured outputs and tool call arguments.

The trick is that nothing gets added to the text. No hidden characters, no metadata: the model just nudges which tokens it picks as it writes. Later the company runs a statistical check and says whether a text is watermarked, not watermarked or uncertain. Longer text gives a stronger signal. The engineer who built it says it runs on almost every token. According to the docs, quality and speed don't change.

You can't check someone else's text yourself: access to the detector is granted only by application. You can check your own responses, though, through an x-watermark header that reads true.

The reason is the EU AI Act, which requires machine-readable labeling of generated content. Opting out is possible, but only for team admins in the console, model by model, and you have to confirm that a legal exception applies and that the responsibility is yours. Every change lands in the audit log, and some models may be locked on.

Grok isn't first here: a few days ago Anthropic announced the same kind of invisible marks for Claude starting November 3. In the replies, some users threaten to cancel, others praise the fact that you can at least turn it off.

Bottom line: AI text is about to give itself away, just not to you.