xAI plans to add invisible watermarks by default to Grok API text output starting November 7
According to xAI’s official developer documentation, the company plans to add invisible watermarks by default to the API text output of the reasoning, non-reasoning, and multi-agent versions of models such as grok-4.7, grok-4.6, grok-4.5, grok-4.3, and grok-4.20 starting November 7, 2026. The watermark applies to global endpoints, as well as ordinary text, structured output, and tool call parameters produced by the API and the Chat Completions API.
The document says the watermark is implemented by making slight adjustments to token selection when the model generates text, without adding any visible characters, marks, or hidden text, with the aim of enabling xAI to identify model-generated content later in order to comply with the EU AI Act requirements. Team administrators can turn off this feature by model in the SpaceXSI Console (some models may have the setting locked on); when turning it off, they must confirm that the applicable legal exception applies. The change takes effect for new requests within about 5 minutes and is recorded in the audit log. The x-watermark: true header in the response can indicate that the output has been watermarked. The document states that this feature does not affect response quality or speed, detection is statistical, and access to the detection API must be applied for.
Original sources:
Title: Output Watermarking Publisher: SpaceXSI Docs (xAI) Date published: 2026-10-31 (page last updated) URL: https://docs.x.ai/developers/advanced-api-usage/watermarking Source nature: official document Verification status: full text read in this review, verification date 2026-11-01 Supported content: list of models with default-on status starting November 7, scope of coverage (ordinary text/structured output/tool call parameters), applicable endpoints, invisible mechanism, team-admin model-by-model opt-out process and conditions, EU AI Act compliance purpose, x-watermark header and the statement that quality/speed are unaffected
ABAB AI Insight
This move by xAI (named as SpaceXSI in the document) comes after it signed only the safety and security sections of the EU AI Act voluntary Code of Practice in 2025 and explicitly avoided the transparency and copyright sections. Article 50(2) requires providers of systems that generate synthetic text and similar content to ensure outputs are marked in machine-readable format and detectable as AI-generated or manipulated; for systems placed on the market before 2026-08-02, the marking obligation has a grace period until 2026-12-02. The official document directly frames the watermark as a way to satisfy that requirement, covering APIs rather than only consumer products.
In terms of capital and business strategy, this is not a financing or M&A event, but a default marking adjustment on the API text generation path: developers do not need to change request parameters, and the output format is basically unchanged (with only a new header added), but the default changes from no watermark to a watermark unless the team administrator actively opts out and takes responsibility for confirming the legal exception. Detection is not public, must be applied for, and the response header provides a direct signal for a single request.
Compared with OpenAI, OpenAI announced on October 5, 2026 that its textGrain watermark would be opt-in for API customers globally (off by default), and would only be planned to turn on by default in the EU for ChatGPT and Codex. xAI, by contrast, chose to enable it by default for the listed Grok API models and allow opt-out by model. Both use statistical token adjustments and do not add visible characters, but their default directions are opposite: OpenAI leaves the decision to API customers, while xAI leaves it to team administrators who must actively turn it off.
This change shifts the default output state of the Grok API to watermarked, requiring developers who need unmarked text to complete the opt-out process and confirm that an exception applies, otherwise new requests will carry the statistical signal. For scenarios that rely on the Grok API to generate content at scale, this makes it easier to identify the model source, while leaving compliance responsibility more clearly with the party choosing to disable it. Consumer-facing products are not required within the scope of the document to turn on the feature by default at the same time.
ABAB News · Cognitive Law
Default API watermarking shifts the burden of identification to all users up front, while leaving the burden of disabling it to administrators who actively confirm an exception.
Statistical token adjustment can mark text without changing visible content, but detection strength still depends on text length and subsequent application permissions.