xAI enables invisible watermarks for Grok models to meet EU rules

Compliance with the EU AI Act

xAI has officially activated invisible text watermarks for seven of its Grok models. This move was designed to satisfy Article 50 of the EU AI Act, a set of regulations governing artificial intelligence. As reported by BornCity, xAI is currently the first major API provider to turn on this type of marking by default.

The new measures apply to several models, including grok-4.7, grok-4.6, grok-4.5, and grok-4.3. It also covers specific variants like grok-4.20-reasoning and grok-4.20-multi-agent. However, the update does not include the grok-build-0.1 development model or older systems like grok-3.

The company is acting ahead of strict enforcement deadlines. While the EU AI Act began taking effect on August 2, 2026, certain systems placed on the market before that date must be marked by December 2, 2026. Failure to comply could result in heavy penalties of up to 15 million Euro or three percent of a company's global annual turnover.

How the invisible watermark works

The watermarking technology functions through a statistical preference for specific tokens during the sampling process. Tokens are the basic building blocks of text that AI models process. By choosing certain tokens from a rotating list more frequently, the system embeds a signal into the output without changing the meaning of the text.

Because the method is statistical, there are no visible characters or formatting marks added to the response. This ensures that the output remains clean for the end user. The signal is designed to survive simple copy-paste actions and the removal of Unicode characters, making it harder for users to strip away the AI's origin.

For a reliable identification, the statistical signal requires approximately 250 tokens of text length. Shorter responses, such as one-liners or small JSON structures, often result in an "undetermined" result during verification. Furthermore, manual rewriting or re-generating the text can weaken the detectability of the watermark.

Impact on developers and users

For developers using the API, this implementation is designed to be seamless. xAI states that no code changes are required because the response format remains unchanged. The company also claims that execution speed and output quality are not affected by the addition of these watermarks, though independent verification of these claims is not yet available.

The scope of the update includes standard text, structured JSON outputs, and arguments in function calls across all endpoints. This applies to both the Response and Chat-Completions interfaces. It covers the global address api.x.ai as well as regional network nodes, ensuring a consistent standard for all API users.

It is important to note that these watermarks are not currently included in consumer-facing offerings. This means that grok.com, associated consumer apps, and integrations within the X platform are not covered by this specific implementation. The focus remains on the professional and developer-facing API tools.

Rules for deactivation and enforcement

The rules for deactivating these watermarks are very restrictive to ensure regulatory compliance. An opt-out is only available to team administrators in the console. If granted, the opt-out applies to the entire team for all API keys and takes effect within approximately five minutes.

To request a deactivation, administrators must confirm a specific EU legal exception. This process is logged by the system, and some individual models may remain locked regardless of the request. This ensures that the company maintains a clear audit trail for why watermarks might be disabled in certain contexts.

Authorities are expected to begin enforcement actions against non-compliant providers starting in the first quarter of 2027. While many other organizations signed a voluntary EU code of conduct for transparency, xAI did not join that specific group. However, they are now moving toward the mandatory requirements set by the law.