The thorny issue of watermarking AI-generated text took an interesting turn this past week, with SpaceXSI seemingly U-turning on plans to automatically watermark text generated by its Grok AI models.
As per its updated output watermarking documentation, Elon Musk’s AI company — which recently rebranded from SpaceXAI — says the feature is switched off by default, with developers able to enable it for individual models through their team’s administration settings.
News of the watermarking feature itself started to emerge in late October, though there was no formal public announcement. The earlier version of the documentation, captured for posterity by the Internet Archive’s Wayback Machine on both November 1 and November 6, was quite different to the existing one, noting that watermarking would be on by default starting from November 7.
The pending change was spotted by some high-profile X users on October 31.
SpaceXSI has since reversed course, however, making watermarking an opt-in feature for developers using Grok through its API. One online documentation archiving service, Spybara, indicates that the change happened some time between Monday night (November 9) and Tuesday morning (November 10), suggesting that the documentation continued to describe watermarking as enabled by default for nearly three days after its scheduled November 7 rollout.
“SpaceXSI has since reversed course, however, making watermarking an opt-in feature for developers using Grok through its API.”
The company has given no public explanation for the change, and it’s unclear whether the original default-on policy ever fully came into effect.
Meeting the watermarking obligations
To rewind just a little, text watermarking is a technique for embedding an invisible, machine-readable signal into AI-generated text. The process subtly influences which words or tokens a model chooses when generating a response, and over a sufficiently long passage of text, those choices create a statistical pattern that can be detected later, helping to establish whether it originated from an AI model.
Detection typically requires access to specialised tools, with the major model labs restricting access to their watermark detectors to approved applicants.
The technology has become increasingly relevant in Europe, where the EU AI Act introduced transparency obligations for providers of AI systems that generate synthetic content. Under Article 50(2), providers must ensure that AI-generated text, images, audio and video are marked in a machine-readable format, with the relevant provisions taking effect on August 2, 2026. A subsequent amendment introduced a limited grace period until December 2 for systems already on the market before August 2, meaning that the clock is ticking for AI companies yet to meet the requirements.
It’s worth noting that the various AI companies have taken different approaches to their text watermarking obligations. In August, Anthropic became one of the first major AI labs to announce text watermarking, applying the feature globally across supported Claude models and products, including its API. The decision proved somewhat controversial, particularly among developers concerned about how watermarking might affect generated code, and the fact that Anthropic provided no mechanism for opting out.
Meanwhile, OpenAI followed suit in October, introducing its own text watermarking system called textGrain. However, it took a different approach, making the feature optional for API developers while introducing automatic watermarking for eligible ChatGPT and Codex text output in the EU.
Google, for its part, has been deploying its own SynthID watermarking technology across various products and services over the past couple of years, including its Gemini API. The technology supports text alongside images, audio and video, with the text watermark embedded into the model’s output as it generates responses. In July, the company confirmed it was signing the EU’s Code of Practice on Transparency of AI-Generated Content.
Notably, SpaceXSI’s updated documentation also confirms that it’s using SynthID Text for its own watermarking feature.
A change of heart
SpaceXSI was the only major AI lab that hadn’t signed the EU’s Code of Practice on Transparency of AI-Generated Content, with at least one report at the time suggesting it would steer clear of watermarking Grok’s text.
However, the emergence of SpaceXSI’s documentation in late October was the first clear indication that it was preparing to join the fray. And the original text was quite explicit about the company’s rationale. “SpaceXSI watermarks text model output to comply with the EU AI Act,” the document noted.
“Watermarking can help enable compliance with the EU AI Act.”
That statement has since been replaced by more cautious language. “Watermarking can help enable compliance with the EU AI Act,” it now reads.
That shift is notable, insofar as the original wording reflected a policy in which SpaceXSI took it upon itself to automatically watermark Grok’s output, ostensibly to comply with the EU AI Act. Now, the company appears to be passing the responsibility for enabling watermarking to developers, while being more circumspect about whether the technology is sufficient to satisfy the bloc’s legal requirements.
There are some practical consequences to all this, too. The feature is controlled at team level, meaning an administrator can enable or disable watermarking for individual Grok models, with the setting applying to every API key belonging to that team. There is no option to switch watermarking on or off for individual requests, meaning developers serving customers in different regions can’t apply different watermarking policies to the same model within a single team. They would need to maintain separate team configurations to do so.
The company’s revised documentation establishes that watermarking is available across seven Grok models, including Grok 4.7 and 4.6, but disabled by default. What remains unclear is why SpaceXSI abandoned its previously stated plans to make the feature automatic.
The New Stack has reached out to SpaceXSI for comment, and will update when, or if, we hear back.
The post SpaceXSI quietly flips Grok API text watermarking to off by default appeared first on The New Stack.
