Astana, 4 November, . From 7 November, responses from Grok models that developers receive through the API will carry an invisible watermark marking the text as generated by artificial intelligence, the developer company SpaceXSI said.
The company explains the change by the requirements of the EU Artificial Intelligence Act. The mark will be applied to responses from seven models, including grok-4.7 and grok-4.20, on all API endpoints, global and regional: plain text, structured output and tool-call arguments. The watermark adds no characters, markup or hidden text to the response: during generation the model slightly shifts its choice of words. It is detected statistically, and the longer the text, the more reliable the result.
What it means: The requirement to label AI texts is ceasing to be a declaration: a major developer is building the mark into the text itself, and it cannot be removed simply by deleting a label.
Who benefits: Platforms and newsrooms that need to tell AI texts apart, and regulators that require labelling; developers who pass off generated text as their own will find it harder; the company has not yet said how the mark can be checked.
Background: Kazakhstan’s law on artificial intelligence also requires synthetic content to be labelled, both with a visible notice and in machine-readable form.
