xAI will switch on an invisible watermark for text produced through its Grok API on November 7, 2026, the company says in its developer documentation. From that day, responses from seven Grok models will carry the mark by default, on every API endpoint, with no change needed to developers’ requests. Team administrators will be able to switch the mark off model by model, but only after confirming that a legal exception covers their use. The change applies to the API only, not to Grok’s consumer apps or its presence on X.
What changes on November 7
The documentation names seven models for the November 7 start: grok-4.7, grok-4.6, grok-4.5 and grok-4.3, plus three grok-4.20 variants for reasoning, non-reasoning and multi-agent work. The list covers the current Grok 4 line rather than older generations, and the company frames the date as a default switch-on, not a new model launch.
The mark will apply on every text generation endpoint, including the global api.x.ai address and regional endpoints, and across both the Responses API and the Chat Completions API. It covers ordinary text replies, structured outputs and the arguments a model supplies for tool calls, so the signal reaches further than plain chat answers alone.
How the invisible mark works
The watermark itself cannot be seen. While the model writes, it slightly shapes which tokens it picks, so no characters, tags or hidden text are added to the reply. The only visible trace in a single response is a header: x-watermark: true means that request’s output carries the mark, while any other value, or no header, means it does not.
Checking a passage later works differently. SpaceXSI scores text against the watermark and reports watermarked, not watermarked or uncertain, with longer text giving a stronger signal. Detection access is granted by application only, so receiving marked output and testing a passage are separate forms of access. Other vendors use statistical enterprise text watermarking too, but access remains the pinch point.
The opt-out, and its limits
Team administrators control the setting for their team, model by model, from a Model output watermarking row in the console’s Team settings. Clearing a model’s checkbox and saving turns its mark off, with changes reaching new requests within five minutes. Every opt-in and opt-out is recorded in the team’s audit log, which administrators can search by event, description or user.
Opting out is not a casual toggle. The console asks the administrator to confirm that an exception under applicable law applies to their use, and that the team bears responsibility for marking or disclosure duty. Some models may also be locked on for a team, appearing in the dialog without a checkbox, with no console route to switch them off.
Original analysis
This change arrived quietly. There is no launch blog post or announcement thread behind it, only a documentation page updated on October 31 and spotted by outside trackers within a day. That fits a compliance deadline more than a product showcase: the EU AI Act’s transparency rules took effect on August 2, and providers have added machine-readable marks ever since.
Against its rivals, xAI is neither first nor last. Anthropic already marks Claude text and is extending the mark to four older models from November 3, while Google keeps its Gemini 4 Carbon model restricted to vetted defenders. xAI’s version sits between those poles: open to every API customer by default, with detection behind an application form.
That split is the change’s weak point. A watermark nobody outside the company can readily check does little for a teacher, editor or voter trying to verify a passage, and a statistical verdict of uncertain will not settle arguments. The x-watermark header helps developers audit their requests, but it says nothing about text copied, paraphrased or passed through another system.
The opt-out design also shifts risk onto customers. Teams that switch the mark off must attest a legal exception, and the audit log means the decision is on record. For most developers, leaving the default alone is the rational move, which is the point: compliance by inertia, with the burden of justifying an exception placed on whoever departs from it.
Why It Matters
For developers, November 7 is a deadline to check a setting, not to rewrite code. Requests need no new parameter and responses keep their shape apart from one header, so the practical work is deciding, team by team, whether any model should be opted out before the default lands. Teams that miss the date will simply start receiving marked text.
Beyond developers, provenance is becoming infrastructure. As AI text spreads through reports and messages, a machine-readable origin signal is one of the tools that can separate generated text from human writing at scale. Its value depends on detection being accessible, a lesson Europe’s deepfake safeguards also illustrate, and on that front xAI’s application-only gate keeps the public at arm’s length.
Background
The EU AI Act’s transparency rules, in force since August 2, 2026, require providers of generative AI to mark synthetic content in a machine-readable way. Anthropic began marking newer Claude models at launch and is now extending the system backwards, while Google limits its newest cyber-defence model to vetted teams, leaving API-wide defaults like xAI’s as the emerging middle path.
Source: xAI’s watermarking docs and coverage of the docs change.
0 Comments