xAI Turns On Invisible Grok Watermark for Seven Models

xAI, now operating as SpaceXSI, officially launched invisible statistical watermarking for seven Grok API models on November 7, 2026. The change applies by default to grok-4.7, grok-4.6, grok-4.5, grok-4.3, and three grok-4.20 variants, marking every API response with a machine-readable signal that nudges token selection without adding visible characters. The company cites EU AI Act Article 50 as its legal basis, with a December 2026 compliance deadline for systems placed on the market before August 2, 2026.

The watermark appears across all SpaceXSI API endpoints, covering the Responses API and Chat Completions API. It applies to ordinary text, structured JSON outputs, and the arguments models generate when calling tools. Developers see no schema changes and no extra parameters, only a new x-watermark: true header in every response that carries the mark.

How the Watermark Works

The watermark operates during generation by subtly biasing which tokens the model selects. No characters, markup, or hidden text are added to the output string. A passage copied from a watermarked Grok answer looks identical to any other text, with no zero-width characters or detectable artifacts that text sanitizers would flag.

Detection relies on statistical analysis. SpaceXSI says its detector returns one of three verdicts: watermarked, not watermarked, or uncertain. Longer passages yield stronger signals because the token-selection bias becomes measurable only across many tokens. Short replies or brief JSON objects often fall into the uncertain band. The company has not published accuracy figures, false-positive rates, or results after paraphrasing or translation, which are the typical weak points of statistical text watermarks.

Access to the detection API is restricted. Teams must apply through a form, and SpaceXSI has not disclosed who qualifies, whether public detection will be offered, or whether it will carry a price. For now, only SpaceXSI and organizations it approves can verify whether text came from a watermarked Grok model.

Seven Models Covered

The watermark applies to the full catalog of current Grok API text models. The list includes grok-4.7, grok-4.6, grok-4.5, grok-4.3, and the three grok-4.20 variants: reasoning, non-reasoning, and multi-agent. The documentation refers to the grok-4.20 models without the "-0309" date suffix used on the official models page. API pricing remains unchanged by the watermarking policy.

The coding model grok-build-0.1 appears on the models page but is absent from the watermarking list. The documentation does not explain whether it is exempt or simply not yet included. The policy applies to text generation only and does not extend to Grok's image, video, or voice products.

Opt-Out Procedure

Team administrators can disable watermarking per model in the SpaceXSI Console, but only after confirming that a legal exception applies and accepting sole responsibility for any remaining marking or disclosure obligations. The opt-out flow requires opening Team settings, where the "Model output watermarking" row shows On (all models), Off (all models), or Custom status. Each model has its own checkbox.

Turning a model off requires confirming with an "Opt Out" button. New requests reflect the change within five minutes. Every opt-in and opt-out is recorded in the team audit log. The setting is team-wide and applies to every API key in the team, so companies that need different behavior for customer-facing and internal workloads must run them under separate teams. Some models may be locked on for a given team; those appear in the dialog without a checkbox, and affected users must contact their team admin or support.

EU AI Act Compliance

SpaceXSI cites the transparency chapter of the EU AI Act as its legal basis. According to the European Commission's published summary, providers of generative AI systems must apply a machine-readable mark to synthetic content and make that mark detectable. The obligation covers text as well as images, video, and audio, and it applies from August 2, 2026.

The Commission lists exceptions. Marking is not required when an AI system performs an assistive function for standard editing or when it does not substantially alter the input data supplied by the deployer. Those exceptions explain the shape of the opt-out in the console, where an administrator must state that one of them applies.

The Commission also describes a grace period. For generative AI systems placed on the market before August 2, 2026, the marking obligation under Article 50(2) applies only from December 2026. A default switch on November 7 lands about a month before that grace period ends. Fines for breaching the transparency obligations can reach €15 million or 3% of worldwide annual turnover.

Release Date

The watermark officially launched on November 7, 2026. The change is described on a new "Output Watermarking" page in the official developer documentation, last updated on October 31, 2026. SpaceXSI documentation states that the watermark is enabled on every API endpoint, including the global endpoint at api.x.ai and the regional endpoints.

Impact on Developers

For most applications, nothing breaks on November 7. Requests stay the same, the response schema stays the same, and the content of answers does not gain any extra characters. Applications that log responses can record the x-watermark header alongside each completion, giving an audit trail of which outputs were marked.

Agent builders have the most to test. Because the watermark applies to structured outputs and tool-call arguments, it touches the parts of a response that software parses rather than people read. SpaceXSI claims quality is unaffected, but teams running JSON-heavy workflows or long chains of function calls should compare schema-validation failures and tool-selection errors before and after November 7, at least until independent measurements exist.

What This Means

SpaceXSI's default watermark places the burden of compliance on the company rather than on individual developers. Teams that publish Grok-generated text to EU end users will generally want the watermark on, since machine-readable marking is exactly what the AI Act requires. The undisclosed detection accuracy and restricted access to the verification API leave gaps in the transparency promise, but the logged opt-out procedure and per-model controls give enterprises the audit trail they need. The December 2026 grace-period deadline means this default will become the norm for generative AI providers operating in Europe.