Analysis
xAI now watermarks text from seven Grok API models by default, citing the EU AI Act, but the mark is statistical, short replies are hard to detect and no accuracy figures are published
From November 7, xAI watermarks text from seven Grok API models by default, citing the EU AI Act. The mark is statistical and no accuracy rates are published.
xAI began applying an invisible watermark by default to text returned by seven Grok models through its API on November 7, 2026, saying it does so to comply with the EU AI Act, per NextFor.ai and Data Studios. The framing needs a limit: the documentation describes the API, not every Grok surface, the mark is a statistical pattern in word choice rather than hidden characters, and short outputs often cannot be classified. xAI has not published detector accuracy or false-positive rates.
The essentials
- What changed: from November 7, text from grok-4.7, 4.6, 4.5, 4.3 and three grok-4.20 variants carries a watermark by default. The xAI documentation page is dated October 31, per NextFor.ai.
- How it works: the mark slightly shifts token choices during generation and adds no characters, so copying, reformatting or stripping invisible Unicode does not remove it, per NextFor.ai.
- Scope: it applies to all endpoints, including the global and regional ones, across the Responses and Chat Completions APIs, and covers text, structured outputs and tool-call arguments, per NextFor.ai and Data Studios.
- Opt-out: team admins can disable it per model, but must confirm a legal exemption applies, and remaining labeling duties fall on them, per NextFor.ai.
What a developer sees and can check
Responses include an x-watermark: true header, which exists only in the HTTP response and does not travel with copied text, per NextFor.ai. The detector returns watermarked, not watermarked or uncertain, and short replies such as one-line answers or small JSON objects often land in uncertain. Access to the detector requires applying through a form. xAI has not published accuracy or false-positive rates, nor results for text edited or translated after generation.
The EU AI Act framing
xAI's documentation ties the change to the transparency rules of the EU AI Act, per SuperGrok, which refers to Article 50. ToolNavs and Data Studios say teams that publish Grok text to users in the EU will generally want the watermark on, because turning it off does not remove the underlying disclosure obligation. xAI says the mark does not affect quality or speed, a company claim without supporting figures.
What is not established
- Detector accuracy and false-positive rates: not published.
- Whether the mark survives editing or translation: no results published.
- Whether the Grok consumer app or other surfaces are watermarked: not covered by the sources we read.
- Whether xAI signed the EU's voluntary code of practice: not addressed in the sources we read.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Start date | November 7, 2026, on by default | NextFor.ai |
| Models | grok-4.7, 4.6, 4.5, 4.3 and three grok-4.20 variants | NextFor.ai |
| Coverage | Text, structured outputs and tool-call arguments | Data Studios |
| Opt-out | Team admins per model, asserting a legal exemption | ToolNavs |
| Stated reason | Compliance with EU AI Act transparency rules | SuperGrok |
FAQ
Does every Grok answer carry the watermark?
Not every surface. The documentation covers text returned by seven models through the API; the sources we read do not cover the consumer app.
Can the watermark be removed by copying or reformatting?
According to NextFor.ai, no: it sits in word choice, so there is no hidden character to strip. Heavy editing or translation has not been tested publicly.
Can I turn it off?
Team admins can disable it per model by confirming a legal exemption applies. The disclosure duties then fall on the team.
Method and limits
Sources checked on November 8, 2026. We did not read xAI's documentation directly; we relied on NextFor.ai, Data Studios, ToolNavs and SuperGrok, which summarize it. xAI's quality and speed claims are unverified. This is not legal advice.
Sources
- NextFor.ai: Grok API Invisible Watermark: What Changes on November 7 (verified 2026-11-08)
- Data Studios: Grok API to watermark text from seven models by default from November 7 (EU AI Act) (verified 2026-11-08)
- ToolNavs: Grok API text will carry a watermark by default from November 7, with per-model opt-out for team admins (verified 2026-11-08)
- SuperGrok: Grok API Text Watermarking Explained (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy