SpaceX will watermark Grok text by default to meet EU AI Act rules

Meeting European regulations

SpaceXSI, the AI unit for SpaceX, is preparing for new regulatory requirements in Europe. The company plans to add an invisible watermark to text generated by seven different Grok models on its API. This change is specifically designed to comply with the EU AI Act, which governs how artificial intelligence systems are deployed and managed.

The timeline for this rollout is quite specific. The watermarking will be enabled by default for every team starting on November 7, 2026. This date falls just 25 days before the official December 2, 2026, deadline. That deadline applies to providers of AI systems that were already on the market before August 2, 2026.

According to the AI Act Explorer, the rules require providers of synthetic text to ensure their outputs are marked in a machine-readable format. The goal is to make the content detectable as being artificially generated or manipulated. By implementing this now, SpaceXSI is positioning itself to meet these legal standards ahead of the final deadline.

How the invisible watermark works

The watermarking system used by SpaceXSI is designed to be completely invisible to the average reader. The documentation explains that the watermark slightly shapes which tokens the model picks during generation. Because of this method, the system adds no extra characters, markup, or hidden text to the final response provided to the user.

Developers interacting with the API will not see a specific request parameter to toggle this feature. Instead, a response header labeled "x-watermark: true" will indicate whether the output contains a watermark. This allows the underlying technology to remain seamless for the end user while still providing a signal for those who need to verify the source.

The detection process is statistical rather than a simple binary check. SpaceXSI will score a piece of text against the watermark to report if it is watermarked, not watermarked, or uncertain. The documentation notes that longer pieces of text provide a stronger signal for this detection method. Access to the specific detection API is currently granted on an application basis.

Provider defaults vs. opt-ins

SpaceXSI has chosen a different path for compliance than some of its larger rivals. While OpenAI announced a text watermarking feature in October 2026, that system stays off by default unless customers choose to opt in. In contrast, SpaceXSI is making the watermark the standard experience for all teams using their API.

This means that most customers will receive watermarked content automatically without having to change any settings. If a team wants to disable the watermark, they must manually switch it off in the SpaceXSI Console. This action applies to every API key held by that specific team and is recorded in the team's audit log for transparency.

The company is placing the legal weight on the customer who chooses to opt out. To turn off the watermark, an admin must confirm that a legal exception under applicable law applies to their specific use case. By doing so, the customer agrees that they are solely responsible for any marking, disclosure, or other legal requirements that might arise from using unwatermarked content.

Impact on developers and labs

For developers building applications on the SpaceXSI API, this change requires very little technical modification. Since there is no request parameter to manage, the watermarking happens automatically in the background. Developers can simply check the response headers to see if the watermark is active on the generated text.

The move also highlights a growing trend in how AI labs handle governance and trust. By making watermarking a default, SpaceXSI is taking on a portion of the governance work for its customers. This helps businesses navigate the complex web of international AI regulations without having to build their own detection tools from scratch.

While this is a significant move for the Grok models, it is not yet clear how robust the watermarks are against human editing or paraphrasing. SpaceXSI has not provided details on how well the watermark holds up when a user modifies the AI's output. However, the move signals a clear commitment to following the machine-readable marking rules established by the EU.