xAI Turns On Hidden Text Watermarking Across Seven Grok Models

Summary

• Grok API text outputs carry a hidden statistical watermark by default starting 7 November 2026.
• Seven models are covered; only team admins can switch it off, per xAI documentation.
• Opting out requires accepting sole legal responsibility for any required labeling or disclosure.

The latest

Text generated through the Grok API now carries an invisible statistical watermark by default, applied during generation by subtly steering token selection, according to the Output Watermarking page in xAI's developer documentation. Seven models are covered across all API endpoints, with no request parameter and no code change required. xAI cites compliance with the EU AI Act as the stated reason.

Details

• The models: Coverage spans 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, per the documentation, last updated 31 October 2026. Independent write-ups noted that grok-build-0.1 does not appear on the watermarking list.
• The scope: The mark applies to the global api.x.ai endpoint and regional endpoints, through both the Responses API and Chat Completions API, and covers plain text, structured outputs and tool-call arguments. Image, video and audio products are excluded, according to xAI.
• The mechanism: xAI says the watermark is a statistical signal embedded by slightly influencing token choice, adding no characters, no encoding and no hidden text. The company states it does not affect response quality or latency, a claim published without supporting measurements.
• How to verify: The x-watermark response header returns true when an output is marked; any other value, or a missing header, means unmarked. Separate text can be checked through a detection interface granted on request via a signup form linked from the documentation.
• The verdicts: Detection returns watermarked, not watermarked, or uncertain. Longer texts produce a stronger signal, while short replies may land in uncertain, per xAI. Accuracy rates, false-positive rates, and results under paraphrasing or translation have not been published.
• The opt-out: Only team admins can change the setting, from xAI Console team settings, where a Model output watermarking row offers On, Off or Custom with per-model checkboxes. The setting applies to the entire team, not a single key, and some models may be locked On for certain teams, appearing without a checkbox.
• The condition: Disabling a model requires an Opt Out confirmation, an acknowledgment that a legal exemption applies, and acceptance of being "solely responsible" for any required labeling or disclosure. Changes take effect for new requests within 5 minutes and are recorded in the team audit log.
• The regulation: A Data Studios report dated 1 November 2026 said Article 50(2) of the EU AI Act, as amended by the AI Omnibus package, requires synthetic text to be marked in machine-readable form. The rules apply from 2 August 2026, with December 2026 the deadline for generative systems released earlier.
• The penalties: Fines under the framework reach 15 million euros or 3% of global annual revenue, according to the same report. xAI's documentation names EU AI Act compliance as the reason for the default, without detailing its own exposure assessment.
• Open questions: Independent reports published between 31 October and 7 November noted the documentation does not state whether consumer Grok apps carry the same mark, and that the x-watermark header belongs to the HTTP response and does not travel with copied text.

Between the lines

Two design choices shift the burden onto customers. Because the opt-out sits at team level and requires accepting sole responsibility for labeling, switching it off converts a vendor-side compliance default into a customer-side legal exposure. And because the verification header does not travel with copied text, downstream recipients cannot confirm origin without the separate detection interface xAI grants on request.

What's next

The December 2026 deadline for generative systems released before 2 August 2026 is the next fixed marker. Also watch whether xAI publishes detection accuracy and false-positive rates, and whether consumer Grok apps are brought into scope.

Source: xAI developer documentation, Data Studios, explainx, NextFor AI, AIToolRanked, Superpower Daily