xAI's Grok API Starts Watermarking Every Text Response Today Across Seven Models, Flagged by an x-watermark Header

xAI's text watermarking policy for its developer API took effect on November 7, 2026, the date the company's own documentation has carried since a guide update published on October 31. From that date, text generated through the API by seven supported Grok models carries an invisible watermark by default: grok-4.7, grok-4.6, grok-4.5, grok-4.3, and the three variants of the grok-4.20 family — reasoning, non-reasoning and multi-agent.

The mechanics matter for developers because nothing visible changes. According to the documentation, the watermark is a statistical pattern in the model's choice of tokens rather than a string of hidden characters appended to the output. There is no request-code change required to receive it, and coverage extends beyond plain prose to structured outputs and even the arguments a model passes to tools. The provider marks whether a given response was watermarked through a response header: x-watermark: true means the marking was applied, while another value or its absence means it was not.

xAI's stated rationale is compliance with the EU AI Act's transparency obligations, which require that AI-generated content be identifiable as such. The company is not alone in moving on this timeline — Anthropic has been retrofitting invisible text watermarks onto older Claude models in waves to meet the same rules, with a second group of models scheduled for mid-November, and OpenAI turned on watermarks for ChatGPT text in the EU earlier this year. The difference is that xAI's rollout makes watermarking a default property of API output across an entire model family rather than a regional or consumer-only feature.

There is an escape hatch, but it is administrative rather than per-request. Team administrators can opt out on a per-model basis, and the setting applies across all of a team's API keys. Consumer Grok chats in the app are governed by separate policies; the API documentation does not establish that consumer output carries the same mark. Enterprises that need to keep an audit trail are advised to record the x-watermark status alongside each response, because intermediaries that strip headers leave them with no record either way.

Detection is deliberately gated. xAI offers a detector application, but access is granted per team and requires administrator involvement — there is no public paste-in checker, and the public pricing page does not list a tariff for detection. That asymmetry is worth being clear-eyed about: the vendor and regulators can check for the mark, but the people receiving AI-generated text generally cannot. NIST's published guidance on invisible watermarks notes that such detectors can both falsely find marks that are not present and miss ones that are, and xAI's guide publishes no error rates or minimum text length for reliable detection.

The deeper limitation is logical rather than technical. A missing watermark is not evidence of human authorship — text can be generated without a mark through opted-out teams, other providers, or models with no watermark at all, and heavy human editing can sit on top of watermarked output. What the change actually establishes is a provable record of origin at the point of generation. For businesses that generate customer replies or published copy through Grok's API, the practical to-do list is short: decide whether to keep the default on, record header status with outputs, and treat any later detection claim about copied text with appropriate skepticism.