Grok’s developer guide sets November 7, 2026 for default text watermarking on seven API models, affecting businesses that generate customer replies or published copy through the service. For those businesses, recording whether a response was marked is different from checking a passage someone later copies elsewhere.
The guide was updated on October 31 and gives no activation hour or timezone. November 7 is the scheduled start, not confirmation of a completed rollout.
Which text is covered
The listed models are 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. Coverage includes prose, structured outputs and tool-call arguments—the data a model sends to a tool. No request-code changes are required.
The provider describes a pattern in the model’s choice of text pieces, not extra hidden characters. This API policy does not establish coverage for consumer Grok chats.
A response record is not a text detector
On the original provider response, x-watermark: true means marking was applied; another value or an absent header means it was not, according to the header documentation.
For a business keeping an origin record, save that status with the corresponding output, model and request time. If an intermediary has discarded the headers, record the status as unobserved; missing records are not evidence that the original response lacked a mark.
Checking copied text requires separate access. The linked detector application requires a team administrator and says access is granted per team. An application route is not a public paste-in checker or a promise of immediate approval.
The public pricing page does not list a detector tariff. That omission is not evidence that detection is free. Before relying on the service, teams should establish its price, usage limits and handling of submitted text.
What a result cannot establish
The documented detector can report watermarked, not watermarked or uncertain; longer passages provide stronger signals. The guide supplies no error rates or minimum text length.
The US National Institute of Standards and Technology explains that invisible-watermark detectors can make both kinds of error: finding a mark that is not present, or missing one that is. That is a general limitation, not a measurement of Grok’s accuracy.
Grok also permits some model-level opt-outs under stated conditions. Settings apply across a team’s API keys. The implication is straightforward: AI-generated text can exist without this mark, so a negative result cannot prove human authorship.
Keep a generation record separate from a later detector verdict. An uncertain result should remain uncertain, and having no detector access means no check occurred. A mark concerns a text’s origin, not whether its claims are true or how much human editing followed.
For context on that distinction in another system, see RohitAI’s explanation of Claude’s text watermark and authorship limits. It is not a route to checking Grok text.