Grok API documents switch to opt-in text watermarking

SpaceXSI’s developer documentation now says text watermarking for Grok API output is disabled by default, replacing earlier guidance that supported models would watermark answers automatically from Nov. 7. Team administrators must now enable the feature individually for each model in the SpaceXSI Console.

The change is documented on an Output Watermarking page last updated Nov. 10. The New Stack reported that archived versions of the page on Nov. 1 and Nov. 6 had said text watermarking would be enabled by default starting Nov. 7, and that the documentation appeared to change between the night of Nov. 9 and the following morning.

It remains unclear whether the original default-on policy was ever fully implemented. SpaceXSI has not publicly explained the new default, and the documentation does not establish when the underlying service behaviour changed.

The current system is configured at team level. Once enabled for a model, watermarking applies automatically to new text-generation requests across every API key belonging to that team, without code changes. The setting covers the Responses API and Chat Completions API, including ordinary responses, structured output and tool-call arguments. Changes apply to new requests within five minutes, and every opt-in or opt-out is recorded in the team audit log.

Some models may be locked on for a team and cannot be disabled through the Console. Developers can use the x-watermark response header to check an individual result: a value of true indicates that its output was watermarked, while another value or no header means it was not.

The watermark is not a visible label or added hidden text. SpaceXSI uses Google DeepMind’s SynthID Text, which subtly shapes the model’s token selection as it generates output. Detection is statistical, with results reported as watermarked, not watermarked or uncertain; longer and more varied passages create a stronger signal, while heavy rewriting or translation can reduce detection confidence. Access to the detection API requires an application.

The documentation says watermarking can help enable compliance with the EU AI Act. Article 50(2) requires providers to mark AI-generated text, images, audio and video in a machine-readable format. The relevant provisions were scheduled to take effect Aug. 2, with a limited grace period until Dec. 2 for systems already on the market before that date.

Audio follows different rules. The Voice API applies an inaudible AudioSeal watermark to Text to Speech and Speech to Speech output, including custom voices, when requests use the EU endpoint or end users are in the EU, EEA or California. Those required audio marks cannot be removed; elsewhere, developers can request a watermark through an x-watermark: true header or a WebSocket URL parameter. The company says watermarking has a negligible effect on speed and does not change response quality.

The revised documentation makes text watermarking available across seven Grok models, according to The New Stack, including Grok 4.7 and Grok 4.6.