Grok API adds blind watermarking to text output from Nov 7

On November 3, 2026, Landian reported that SpaceXSI had sent out an email notification stating that Grok API would begin adding a blind watermark to text output by default starting November 7, 2026. Although the feature is being introduced to comply with the EU AI Act, the policy applies to API calls from all regions worldwide, meaning every user generating content through Grok API will receive the watermark.

According to the notice, the blind watermark is formed when the model slightly adjusts Token selection while generating text. No visible characters, invisible characters, markers, or hidden text are added to the output. Detection works through statistical analysis of the text, producing one of three results: watermarked, not watermarked, or uncertain. Accuracy depends directly on text length, since longer AI-generated text carries more statistical signal and is judged more reliably, while shorter or rewritten text is less likely to be detected. The watermark detection API currently requires applying for special access permissions as well.

Users who believe their use case qualifies under a statutory exception can have an administrator turn the watermark off for the relevant model by navigating to console, team settings, general, then model output watermark. This setting applies to all API keys under the team and typically takes effect about 5 minutes after the change. Because every Grok API account is provisioned as a team or organization type, even individual users can disable the mechanism through these settings, provided they are not located in the EU or their product does not target EU users. As of the report's publication, the mechanism had not yet gone live and remained in the notified, pending stage.