Grok API now watermarks text by default ahead of EU AI Act deadline

The models are grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants: reasoning, non-reasoning and multi-agent[1]. Watermarking is on across every API endpoint, including the global api.x.ai endpoint and regional endpoints, and applies to both the Responses API and the Chat Completions API[1]. It covers ordinary text, structured outputs and tool call arguments[1][2].

The watermark is applied during generation by slightly shaping which tokens the model picks, so it adds no characters, markup or hidden text[1]. Detection is statistical: xAI scores a passage and reports it as watermarked, not watermarked or uncertain, with longer text giving a stronger signal[1]. Reporting says access to the detection API is granted by application[2][3].

Developers do not need to change their requests. A new x-watermark header shows whether a given response was marked, with true meaning it was[1].

Watermarking is controlled per model for a whole team, not per key or request, and team admins can opt out in the console[1]. Coverage of the change says opting out also requires confirming a legal exception[2][4].

DeAI notes the change lands about three weeks before the 2 December marking deadline under Article 50(2) of the EU AI Act, and calls this the first confirmed default-on text watermark at a major frontier API, giving an early answer to how providers will meet that deadline[4]. For anyone publishing or reselling Grok output, provenance is now built in unless an admin actively switches it off[1].