Since November 7, 2026, every response from seven of Grok’s models has arrived carrying something invisible: a statistical watermark baked into the text itself, with no characters added and nothing a reader would ever notice. xAI — now folded into SpaceX and recently rebranded SpaceXAI, with owner Elon Musk floating the name SpaceXSI on X — switched the feature on by default across its entire API, a reversal from a company whose own chatbot had publicly resisted doing exactly this just months earlier.
What actually changed on November 7
Per xAI’s own developer documentation, text responses from 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 now include the watermark by default, applied while the model generates its answer. It covers ordinary chat responses, structured outputs, and tool-call arguments, across both the Responses API and Chat Completions API, on every endpoint worldwide — the global `api.x.ai` address and regional endpoints alike. The company states plainly that the purpose is compliance with the EU AI Act.
Developers get one visible signal: an `x-watermark` response header, which reads `true` when a given response was marked. Detecting the watermark after the fact is statistical rather than absolute — xAI’s detector scores text as watermarked, not watermarked, or uncertain, with longer passages giving a clearer read. Access to that detector isn’t public; it’s granted by application. Team admins, and only team admins, can turn the feature off per model inside the console’s Team Settings, but doing so requires checking a box affirming the team is relying on a legal exception and accepts sole responsibility for whatever disclosure law still requires — and the choice is logged in a searchable audit trail.
The irony: xAI was supposed to be the holdout
This is the part that makes the story worth a second look. Article 50 of the EU AI Act, whose transparency obligations took effect August 2, 2026, requires providers to mark AI-generated content in machine-detectable ways “as far as technically feasible.” Anthropic moved first, quietly turning on invisible watermarking for Claude’s text output that same month. OpenAI followed on October 5, confirming it would watermark ChatGPT and Codex output for EU users and open a worldwide opt-in for API customers — notably, opt-in, not on by default, and OpenAI’s own announcement acknowledged it had previously held the feature back over concerns about false positives and competitive risk.
Grok was the conspicuous exception. Coverage at the time, including a widely-shared City AM headline — “ChatGPT to follow Claude’s watermark pledge — but Grok to swerve it” — framed xAI as the lab willing to skip the practice altogether. Five weeks later, it didn’t just adopt watermarking; it shipped the most aggressive version of it in the industry, on by default across seven models rather than requiring anyone to opt in.
Why this happened inside SpaceX, not just xAI
The timing lines up with a corporate reshuffling that’s easy to lose track of. SpaceX acquired xAI outright on February 2, 2026. The combined AI business operated under the xAI name for a few more months before formally rebranding as SpaceXAI on July 6, 2026 — new logo, new handles, same underlying Grok product. On October 4, Musk said on X that the business would be renamed again, to SpaceXSI, with “SI” standing for superintelligence; as of this writing that second rename remains an announcement rather than a completed rollout, though xAI’s own developer docs have already adopted the SpaceXSI name internally. Grok itself isn’t being renamed — for anyone actually using the product, nothing about this corporate churn is visible day to day.
What’s notable is that a company under intensifying regulatory and reputational scrutiny — Grok has drawn EU attention this year over content-moderation controversies — chose compliance-by-default over the minimal opt-in path OpenAI took. For developers building on the Grok API in the EU or anywhere else, the practical upshot is simple: as of November 7, assume every supported model’s output is watermarked unless a team admin has explicitly switched it off, and don’t expect the watermark to be something end users can see, strip out, or rely on as definitive proof of AI authorship — xAI’s own documentation, like OpenAI’s, is careful to note detection is probabilistic, not certain.
For everyday Grok users, nothing looks different. The change lives entirely in the API layer, in a response header most people will never see, enforcing a transparency rule most people have never heard of — which is, in its way, exactly what the EU AI Act was designed to produce.
