xAI starts invisibly watermarking Grok API text by default
These are the company's own documentation and claims, including that watermarking does not change quality or speed; this edition saw no independent test of how reliably the mark survives editing or how often detection errs. Detection is statistical, returns an uncertain result for some text, and is available only by application. The page covers API text output only, not images, audio or the consumer app, and it does not say which legal deadline applies.
What was claimed?
xAI's API documentation, which names the company SpaceXSI, says that starting November 7, 2026, text responses from seven Grok models (grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants) include an invisible watermark by default on every API endpoint. The watermark slightly shapes which tokens the model picks rather than adding characters, and covers ordinary text, structured outputs and tool-call arguments. A response header, x-watermark: true, shows when it was applied. Team admins can opt out per model after confirming that a legal exception applies and that they are solely responsible for any marking, and some models may be locked on.
What does the evidence support?
Documented availability, not performance proof
It is one of the first default-on text watermarks from a major model provider, and xAI says it is doing so to comply with the EU AI Act. If other providers follow, machine-generated text from their APIs could become detectable by the provider, though not necessarily by anyone else.
What are the limits?
These are the company's own documentation and claims, including that watermarking does not change quality or speed; this edition saw no independent test of how reliably the mark survives editing or how often detection errs. Detection is statistical, returns an uncertain result for some text, and is available only by application. The page covers API text output only, not images, audio or the consumer app, and it does not say which legal deadline applies.
Claim or evidence date: 2026-11-07. Last verified: .
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →