xAI's developer documentation, which now uses the name SpaceXSI, says text from the Grok API will carry an invisible watermark by default from November 7, 2026. These answers come from that documentation unless marked otherwise. For the general idea, see LLM text watermarking.
The Basics
Q: When does the Grok API watermark start?
November 7, 2026. It is on by default.
Q: Which models are watermarked?
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.
Q: What output is covered?
Text from the Responses and Chat Completions APIs, including structured outputs and tool-call arguments, on all endpoints, regional ones included.
Q: Why is xAI doing this?
The docs say it is "to comply with the EU AI Act". They don't cite an article. Press coverage ties it to Article 50, which covers marking AI-generated content, and some reports cite a December 2, 2026 deadline for systems already on the market. Those connections are the press's, not xAI's.
How It Works
Q: Does the watermark change the text I get?
Slightly. It is a bias applied to token choices at generation time. The text reads normally and there are no hidden characters or metadata.
Q: How do I know if a response was watermarked?
Check the x-watermark response header. x-watermark: true means it was watermarked. Any other value, or no header, means it wasn't.
Q: Can anyone detect the watermark?
Detection is statistical and returns watermarked, not watermarked or uncertain. Access to xAI's detection service is by application, so a random third party can't check your text with it.
Q: Does editing remove it?
Heavy editing, paraphrasing, translation and very short outputs weaken any statistical text watermark. Light edits usually don't.
Opting Out
Q: Can I turn the watermark off?
Team admins can, per model, in the Console under Team settings, "Model output watermarking". Changes apply to new requests within about five minutes and are logged in the audit log. Some models can be locked on.
Q: What does opting out require?
You have to confirm that a legal exception applies and accept sole responsibility for marking and disclosing AI-generated content yourself.
For Brands
Q: Does this matter if I use Grok to write marketing copy?
Your copy will carry the mark unless your admin opts out. That rarely matters on its own. It matters if you have disclosure duties in the EU, or if platforms start labelling or down-ranking detected AI text. Keep a list of which vendors watermark by default.
Q: Does the watermark affect how Grok describes my brand?
No. It changes how output is marked, not what the model knows. For brand visibility on Grok, see the Grok brand FAQ.