Grok's API Will Watermark Generated Text by Default Starting November 7, with Model-by-Model Opt-Out
Following Anthropic, xAI (referred to as SpaceXSI in its documentation) will invisibly watermark text generated by Grok models via its API starting November 7, 2026, citing compliance with the European AI Act. Unlike Claude, team admins can disable it on a per-model basis, provided they assert a legal exemption.
Anthropic is no longer alone. According to a developer documentation page spotted on October 31, xAI, which now brands its documentation as SpaceXSI, will add an invisible watermark to all text generated by Grok models via its API starting November 7, 2026: grok-4.7, grok-4.6, grok-4.5, grok-4.3, and the three grok-4.20 variants (reasoning, non-reasoning, multi-agent). The marking will be active by default across all endpoints, citing compliance with the European AI Act. Just hours earlier, we reported that Anthropic is expanding its own watermark to four older Claude models starting November 3.
How Grok’s Watermark Works
The mechanism described in the documentation relies on statistical watermarking: during generation, the model subtly biases token selection without adding characters, tags, or hidden text. SpaceXSI claims this affects neither quality nor latency. The marking covers all text returned by the generation APIs (Responses API and Chat Completions), including structured outputs and tool call arguments, which directly impacts agent developers.
Integration requires no changes: there are no parameters to pass. The only addition is an x-watermark: true response header indicating that a request was watermarked.
Detection is statistical: SpaceXSI analyzes text and returns “watermarked,” “unwatermarked,” or “uncertain,” with a stronger signal on longer passages. Access to the detection API is by application only, through a form.
Conditional Opt-Out
This is the main differentiator from the approach described by Anthropic. With SpaceXSI, team administrators can turn off the watermark on a per-model basis in the console for all team keys; the change applies to new requests in about five minutes. However, it is not a simple toggle: disabling the marking requires confirming that a statutory exemption applies to the relevant use case and that the team assumes sole responsibility for any marking or disclosure requirements. Some models may be locked and cannot be opted out of, and every change is recorded in the audit log.
What Remains Unclear
The documentation page addresses only the API. It does not mention whether text produced in the Grok app or on X will be affected; users on X claim consumer subscriptions are exempt, though we have found no official confirmation. As of November 1, we have also found no announcement on xAI’s blog or its official accounts: the documentation remains the sole primary source. Nor does it specify which organizations qualify for detection access or the criteria used to vet them.
Our take
With Anthropic and xAI on board, text watermarking is quickly becoming standard practice for model providers looking to serve the European market, with the AI Act clearly acting as the catalyst. Yet the two approaches diverge on a fundamental point: Anthropic enforces watermarking across the board, whereas xAI allows enterprises to opt out by assuming their own legal liability. While this provides more flexibility for developers, it undermines the watermark’s value as a traceability tool: an unwatermarked text can never definitively prove it was written by a human. And as long as detection remains restricted to select organizations, authors cannot verify for themselves what others might detect in their text.
Sources
- SpaceXSI (xAI) — Output Watermarking, API documentation Primary source
- X — post by Tech Dev Notes (@techdevnotes), October 31, 2026 Secondary source — Earliest spotted report of the documentation page.
- X — post by Manuel Casalta Petitjean (@GloricsLab), October 31, 2026 Secondary source — Claims consumer subscriptions are not watermarked; unconfirmed by xAI.
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
