Grok API Text Watermarks Are Now Off by Default, Docs Say

SpaceXSI’s latest developer documentation now says Grok API text watermarking is off by default, a change from the default-on approach described in its earlier guidance.

That matters for developers who read our October 31 report on Grok’s hidden text watermark. At the time, the company’s documentation said supported models would start marking their answers automatically on November 7, with an option for team administrators to opt out.

The current Output Watermarking page, last updated November 10, instead tells administrators to enable it per model in the Console. It does not explain the changed default or establish when the underlying service behaviour changed.

The setting applies across a team’s API keys, with changes taking effect within five minutes. Some models can still be locked on. Developers can check an individual response through the x-watermark header: true means that output was watermarked.

This is an invisible signal in generated text, not a visible Grok logo. SpaceXSI says its implementation uses Google DeepMind’s SynthID Text technique.

DeepMind’s explanation describes a system that adjusts the probabilities of the tokens a model chooses as it writes. The resulting statistical pattern can help identify AI-generated text without adding a label readers can see.

There are limits. DeepMind says the technique works best on longer, varied responses and is less effective where answers allow little variation. Heavy rewriting or translation can reduce detection confidence, so a watermark is one identification tool rather than a universal AI detector.

Audio has different rules. The updated Grok docs say voice API output is watermarked for the EU endpoint and users in the EU, EEA or California. Those required audio marks cannot be switched off; other requests can opt in.

Grok’s voice business has already moved beyond chat, including Starlink’s support and sales calls. For teams building on these APIs, the distinction between text defaults and regional audio requirements is worth checking before assuming every output follows the same policy.

Should AI-generated text be watermarked automatically, or should developers choose?