As the EU's Artificial Intelligence Act officially takes effect and continues to advance, major AI companies around the world are accelerating their compliance adjustments. Recently, SpaceXAI officially announced that, starting November 7, it will fully introduce invisible watermarking technology into text content generated by its Grok series of large models.

In terms of technical implementation, this invisible watermark is not created by inserting hidden characters invisible to the naked eye into the text or by attaching metadata. Instead, it is cleverly integrated directly into the model's token selection process when generating text. This means the watermark is deeply bound to the text content at the underlying logical level. It is not only highly covert, but also extremely difficult to strip away through ordinary text copying, pasting, or simple modifications.

In terms of coverage, SpaceXAI APIs in all regions worldwide have now officially enabled the watermarking feature, covering multiple Grok models. However, to accommodate the special business needs and deployment flexibility of enterprise users, the team has also preserved the right of choice for organization administrators, allowing them to independently disable the watermarking feature for a specific model based on actual needs.

It is worth noting that industry observers pointed out that this move to strengthen invisible watermarking in text generation is mainly intended to strictly respond to the compliance requirements of the EU's Artificial Intelligence Act regarding the clear labeling of AI-generated content. As SpaceXAI's update is about to take effect, among the world's current mainstream large-model vendors, ChatGPT in North America has become one of the few representatives at this stage that has not yet fully introduced invisible watermarking into text output.