xAI will begin adding invisible watermarks to text generated by seven Grok models through its API from November 7, 2026. The feature will be enabled by default and cover ordinary text responses, structured outputs and tool-call arguments. The company says the change is intended to help identify AI-generated content and comply with the European Union’s AI Act.
The announcement is significant for developers and businesses using Grok to build AI applications, chatbots and automated agents. Unlike visible labels or disclaimers, the watermark will be embedded statistically during text generation without adding characters or changing the appearance of the response. Team administrators will be able to disable the feature for individual models, subject to a legal confirmation process.
Seven Grok Models Will Receive Invisible Watermarks
According to the company’s developer documentation, watermarking will initially apply to seven models:
| Grok model | Watermarking from November 7 |
|---|---|
| grok-4.7 | Enabled by default |
| grok-4.6 | Enabled by default |
| grok-4.5 | Enabled by default |
| grok-4.3 | Enabled by default |
| grok-4.20-reasoning | Enabled by default |
| grok-4.20-non-reasoning | Enabled by default |
| grok-4.20-multi-agent | Enabled by default |
The policy applies across all SpaceXSI API endpoints, including the global API endpoint and regional endpoints. It also covers both the Responses API and Chat Completions API.
The change is not limited to the final text displayed to users. Structured outputs and arguments generated for tool calls will also carry the watermark.
The documentation further states that future models will receive watermarking by default from their launch, although administrators can manage eligible models through the console.
How Does Invisible Text Watermarking Work?
Traditional watermarks are usually visible. They can appear as logos, labels, signatures or statements identifying the creator of an image, document or video.
Text watermarking works differently.
SpaceXSI says its system applies the watermark during the generation process by slightly influencing which tokens the model selects. Tokens are the small units of text that language models use to construct sentences.
This process does not add visible characters, special formatting, labels or hidden text to the generated response. A person reading the output should not see any obvious difference.
Instead, the company uses statistical detection to identify patterns associated with its watermark.
For example, a developer could receive an ordinary paragraph from Grok without any visible indication that it contains a watermark. A compatible detection system could subsequently analyse the text and determine whether it was generated by a watermarked model.
SpaceXSI says longer text samples provide a stronger detection signal. Its detection system can return three possible outcomes: watermarked, not watermarked or uncertain.
Access to the dedicated watermark-detection API is granted through an application process rather than being openly available to everyone.
Developers Can Check the Watermark Through an API Header
One important feature of the implementation is that developers can check whether an individual API response was watermarked.
SpaceXSI has introduced an HTTP response header called x-watermark.
Its value indicates whether watermarking was applied to the generated output:
| Header result | Meaning |
|---|---|
x-watermark: true | The response was generated with a watermark |
| Any other value | The response was not marked |
| Header unavailable | The response was not marked |
This gives developers a way to record watermark status alongside generated content in their applications.
The company says developers do not need to change their existing API requests. There is no additional parameter to enable watermarking, and the response format remains unchanged apart from the new header.
For businesses using Grok in automated workflows, this could make it easier to track which generated outputs carry the watermark.
Why Is xAI Introducing Text Watermarking?
The stated reason is compliance with the European Union’s AI Act.
Article 50(2) of the legislation requires providers of AI systems generating synthetic text, images, audio or video to ensure that their outputs are marked in a machine-readable format and can be detected as artificially generated or manipulated.
The rules are intended to improve transparency around AI-generated content and make it easier to identify synthetic material.
The European Commission says the relevant Article 50 transparency obligations apply from August 2, 2026. However, AI systems already placed on the market before that date have until December 2, 2026, to comply with the specific marking and detection requirement under Article 50(2).
xAI’s November 7 rollout therefore comes ahead of that December deadline.
The distinction matters because the legislation does not simply require companies to add a visible AI-generated label to every response. It establishes a machine-readable marking and detection obligation for providers, subject to technical feasibility and applicable exceptions.
Watermarking is one technical approach to meeting that objective.
Team Administrators Can Opt Out of Watermarking
Although watermarking will be enabled by default, xAI is giving team administrators control over the setting.
Administrators can disable watermarking for selected models through the SpaceXSI Console.
The process involves opening Team settings, locating the Model output watermarking option and selecting the models for which watermarking should be disabled.
The setting operates at the team level, meaning a change applies to every API key associated with that team.
However, opting out is not an unrestricted technical preference. Administrators must confirm that an exception under applicable law applies to their use and accept responsibility for other marking, disclosure or legal requirements.
The company also says some models may be locked on and unavailable for opt-out.
Changes to the setting apply to new API requests within approximately five minutes. The system also records opt-in and opt-out decisions in the team’s audit log.
This approach gives enterprise customers flexibility while making them responsible for determining whether their particular use qualifies for a legal exception.
Will Watermarking Affect Grok’s Speed or Output Quality?
SpaceXSI says the watermarking process will not affect the quality or speed of generated responses.
It also does not require developers to modify their existing integrations or send additional instructions with API requests.
However, the company has not published independent benchmark results establishing the impact across different workloads or text-generation tasks. Its statements about performance should therefore be understood as the company’s own claims.
Another important limitation is that statistical watermark detection is not the same as definitive proof of authorship. The documentation does not establish that a watermark will survive every possible editing, paraphrasing or text-transformation process.
The strength of detection also depends on the amount of text available for analysis.
The Bigger Picture
AI watermarking is becoming an increasingly important part of how technology companies approach transparency and regulatory compliance. As businesses integrate generative AI into customer support, content production, software development and automated workflows, identifying AI-generated material is becoming more relevant.
The move also reflects a shift in the responsibilities of AI model providers. Instead of leaving all identification and disclosure tasks to developers, companies are beginning to incorporate detection mechanisms directly into their generation systems.
However, watermarking alone cannot resolve every question about AI-generated content. Detection reliability, edited text, legitimate exceptions and the responsibilities of businesses publishing AI-generated material will remain important issues.
Looking Ahead
The November 7 rollout will give developers using the seven supported Grok models a new default feature without requiring changes to their API integrations. Businesses will need to review how watermark status is recorded and how generated text moves through their applications, particularly when outputs are passed between automated agents and external services.
The larger test will be whether invisible watermarks provide reliable identification across real-world use cases. As regulatory requirements develop and more AI providers introduce similar systems, machine-readable content identification could become a standard feature of commercial AI platforms.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
