Text watermarking in SpaceXSI's Grok API is off by default, according to SpaceXSI's API documentation, last updated November 9, 2026. Team admins can turn it on per model in the SpaceXSI Console, and the setting applies to every API key on the team. The New Stack reported on November 10, 2026 that an earlier version of the docs said watermarking would be on by default starting November 7.

#What changed?

SpaceXSI's docs now describe watermarking as an admin choice, not an automatic default. "Text watermarking is off by default. Team admins can turn it on per model in the SpaceXSI Console's team settings", the docs say.

The docs describe the feature this way: "Text watermarking embeds an invisible signal in text generated by Grok models, so the text can later be identified as AI-generated." SpaceXSI says it uses SynthID Text, "the text watermarking technique published by Google DeepMind in a Nature paper in 2024." The watermark adds no characters, markup or hidden text; it slightly shapes which tokens the model picks.

The "off by default" status has one caveat. The docs say: "Some models may be locked on for your team. Locked models are listed in the dialog but cannot be turned off from the Console." The page does not say which teams or models are locked.

On regulation, the docs say only: "Watermarking can help enable compliance with the EU AI Act."

#How does the setting work?

Watermarking is a team-level, per-model switch with no per-request control. According to SpaceXSI's docs:

  • Seven models. It covers grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning and grok-4.20-multi-agent.
  • All text APIs. "Watermarking applies to outputs from every text generation API, including the Responses API and the Chat Completions API." That includes ordinary text, structured outputs and tool call arguments.
  • Whole team. "The setting belongs to the team, so it applies to every API key on the team. You choose it separately for each model."
  • No request parameter. Once a model is turned on, the watermark "is automatically applied to all requests, with no further code changes required."
  • Response header. The response format is unchanged apart from the x-watermark header. A value of true means the output was watermarked.
  • Timing and records. "Changes apply to new requests within 5 minutes." Every opt-in and opt-out is recorded in the team's audit log.
  • Quality. The docs say watermarking does not change response quality and its effect on speed is negligible.

In the Console, admins find the Model output watermarking row on the Team settings page, which shows On (all models), Off (all models) or Custom. Only team admins see it. On turning a model off, the docs say: "Turning a model off asks you to confirm that an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure, or other legal requirement."

Detection is separate. According to the docs, "Access to watermark detection is granted by application." Results come back as watermarked, not watermarked or uncertain.

#What did SpaceXSI plan before?

The earlier plan is documented only in The New Stack's report. The New Stack said news of the feature "started to emerge in late October, though there was no formal public announcement." It reported that Wayback Machine captures from November 1 and November 6 showed docs "noting that watermarking would be on by default starting from November 7."

The New Stack also reported that an archiving service indicated the docs changed between the night of November 9 and the morning of November 10, and that the old docs said SpaceXSI watermarked text output to comply with the EU AI Act, wording since replaced by more cautious language. It wrote that "it’s unclear whether the original default-on policy ever fully came into effect." SpaceXSI "has given no public explanation for the change", according to The New Stack, which said it had asked the company for comment.

For context, The New Stack reported that the EU AI Act's Article 50(2) requires providers to mark AI-generated content in a machine-readable format, with provisions taking effect on August 2, 2026, and a grace period until December 2 for systems already on the market. It also reported that SpaceXSI was the only major AI lab that had not signed the EU's Code of Practice on Transparency of AI-Generated Content.

#How does it compare?

Other labs took different approaches, according to The New Stack. It reported that Anthropic applied text watermarking globally across supported Claude models and its API in August, with no opt-out. We covered Anthropic's plan to extend it in Claude's text watermark coming to older models. The New Stack said OpenAI's textGrain, introduced in October, is optional for API developers but automatic for eligible ChatGPT and Codex text in the EU. Google uses SynthID across products including its Gemini API, the outlet added.

Watermarks are one approach to AI-text provenance. Another is Anthropic's cryptographic prototype in FlashProve, which proves which model wrote a reply.

Our analysis: Under the current docs, the decision sits with each team's admin. Because there is no request parameter, The New Stack notes, teams serving customers in different regions "would need to maintain separate team configurations" to apply different policies to the same model. Developers should check the Console row and the x-watermark header rather than assume either state, since some models may be locked on.

#What is still unclear?

SpaceXSI has not said why the default changed, and neither source says whether watermarking was ever applied by default after November 7. The docs do not say which models are locked on for which teams, or which legal exceptions qualify for turning a model off. Neither source covers the consumer Grok apps.