xAI adds invisible text watermarking to Grok API models
Starting November 7, 2026, xAI will implement invisible statistical watermarking across its Grok API models by default, leaving ChatGPT as the only major North American model without this feature.
The watermark will be automatically applied to all Grok API outputs without requiring developers to enable any settings. This affects multiple model versions including grok-4.7, grok-4.6, grok-4.5, grok-4.3, and grok-4.20, across all text generation endpoints including the main api.x.ai endpoint, Responses API, and Chat Completions API.
How the watermark works
Grok's watermark functions by subtly guiding token selection toward a detectable statistical pattern. While individual sentences appear normal, detectors can identify the fingerprint when analyzing larger text samples. xAI assures that this process doesn't affect response length, quality, or speed.
Opt-out options
Team administrators can disable watermarking on a per-model basis through the xAI Console, though this option is limited to specific legal exceptions and compliance requirements. The update doesn't affect consumer-facing products, focusing instead on the API used by businesses and developers.
Policy reversal
With xAI on board, ChatGPT stands as the remaining North American model without text watermarking.
This change marks a significant shift for xAI, which had previously opposed watermarking AI outputs on the grounds that text should remain unaltered. The new approach maintains visible text integrity while adding statistical transparency, likely influenced by the EU AI Act's requirements for generative AI systems.
Related
