Invisible Watermarking in Grok: How xAI is Preparing for the EU AI Act
Starting November 7, 2026, developers using the Grok model API will notice a significant change under the hood. As part of the preparations for strict European regulation, xAI (appearing in their technical documentation as SpaceXSI) has begun embedding invisible statistical watermarking in all textual output generated by its models.
Why is this happening and what does it mean in practice?
The move is intended to comply with the requirements of Article 50 of the EU AI Act, which comes into effect in August 2026. The law requires providers of generative AI systems to embed machine-readable markers in AI-generated content to enable the identification of synthetic content and reduce risks of manipulation or deception.
It is important to understand: This is not a secret tag or hidden text added to your messages. The technology operates at the Tokenization level: the model makes subtle adjustments in word choice (Tokens) during text generation. These statistical patterns allow dedicated detectors to identify with high probability that the content was machine-generated, without compromising the quality of the response or the speed of the model's reaction.
Which models are affected?
The watermark is enabled by default for seven key models, including:
- •The grok-4.7, 4.6, 4.5, 4.3 series
- •grok-4.20 versions (including Reasoning, Non-reasoning, and Multi-agent)
The marking applies to all types of output: free text, structured output (JSON), and Tool call arguments.
How do you check if the output is marked?
For developers, the check is very simple. In every API call (via the Responses API or the Chat Completions API), a response header named
x-watermark- •If the value is
, the output was generated with a watermark.true - •If the value is missing or different, the output is not marked.
In addition, xAI offers a dedicated API for watermark detection, which allows you to input text and receive a statistical rating of the probability that it was generated by the model. The longer the text, the stronger the statistical signal and the more accurate the detection.
Can it be disabled?
Yes, but not for everyone. Team managers can disable the watermark for specific models via the API management console, subject to legal justification. The change in settings affects all API Keys in that team and goes into effect within a few minutes.
Summary and conclusions
xAI's move signals the end of the "Wild West" era of unidentified AI content. For developers building apps based on Grok, it is recommended to review your existing pipeline – especially if you perform byte-to-byte comparisons (such as in Caching or Deduplication), as the change in Tokens might affect the final result.
For more information on the technology and API settings, you can refer to the official documentation: Output Watermarking - xAI Docs EU AI Act - Article 50 Transparency Obligations