Grok Adds Default Text Watermarks
SpaceXSI now watermarks text output from seven Grok API models by default, with opt-out limited to team admins.
Updated on Nov. 8, 2026 in Artificial Intelligence
Live Poll
Should AI companies be required to watermark the text their tools generate?
Beginning November 7, text output from seven Grok API models is invisibly watermarked by default, covering grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants. The watermark applies across SpaceXSI API endpoints and both the Responses and Chat Completions APIs, and developers can opt out only through team admins.
Why it matters
SpaceXSI says the watermarking complies with the EU AI Act, which pushes AI providers toward labeling machine-generated text. Because the mark is embedded in the output itself rather than added as visible content, it can be checked later without changing how the text reads to users.
The watermark slightly shapes which tokens the model selects during generation, and watermarked responses include an x-watermark header set to true. Detection is statistical and returns watermarked, not watermarked, or uncertain, with access to the detection API granted by application.
The players
SpaceXSI
SpaceXSI is the company that operates the Grok AI models and the api.x.ai platform through which developers access them.
The details
Structured outputs and tool call arguments also carry the watermark, and no extra characters, markup, or hidden text are added to responses. Existing integrations require no code changes, and the opt-out setting takes about five minutes to reach new requests. Longer text yields a stronger detection signal, while SpaceXSI says response quality and speed are unchanged.
Timeline
November 7: Default text watermarking takes effect for the Grok API.
November 8: WebProNews publishes its article on the Grok watermarking change.
Industry Shift
The EU AI Act sets the regulatory pressure behind Grok's default watermark, and OpenAI described its own API text watermarking approach in October 2026. Anthropic has said future Claude models will carry watermarks, and Google opened its public SynthID Detector for images, video and audio that same month.
Developers using the Grok API will see watermarked text in outputs from the covered models unless their team admin opts out. Applications that depend on unmarked text will need to account for the watermark, and access to the detection API requires applying for approval.
The takeaway
Developers building on Grok should review whether the default watermark affects their downstream systems before the change reaches their production traffic. Teams that need to detect or opt out of watermarking should coordinate with their admins now, since that control is limited to them.
Further reading
For broader context on how AI providers are approaching output labeling, see the Artificial Intelligence section.
Source note: This article includes information reported by WebProNews.
Live Poll
Should AI companies be required to watermark the text their tools generate?
