← Sunday, November 15 deck

07Policy

Grok Text Watermarking Is Now Off by Default in the SpaceXSI API

Grok's text watermarking is now off by default in the SpaceXSI API. Elon Musk confirmed it and went further. He says text watermarking with Grok is only on where the law requires it or a user asks for it. So if you build on Grok, your output won't carry a hidden watermark unless you switch it on or a law requires one.

Context

Why it matters

Builders using Grok through the API will not have a hidden mark in their text unless they turn it on or a law requires it, which weakens a tool meant to show when text came from a model.

Mark Kretschmann reported on November 15 that Grok's text watermarking is now off by default in the SpaceXSI API, and that Elon Musk confirmed it is only on where the law requires it or a user asks for it. The next morning, Musk quoted that post with the single word Important. SpaceXSI's developer docs now match the default off setting. Kretschmann's post is used as the lead because it carries the factual account; Musk's quote is official backing, but it has no details and landed on November 16.

Deep dive

Kretschmann wrote that Grok's text watermarking is now off by default in the SpaceXSI API, that Musk confirmed the change, and that Musk said the mark is only on where the law requires it or a user asks for it. Musk later quoted the post with the word Important, which is official attention but not a fuller policy statement. The current SpaceXSI developer docs agree that text watermarking is off by default, and that team admins can turn it on for listed Grok models in the console. The docs also say audio from the voice APIs is watermarked by default, except where a team opts a voice model out and the user's location does not require the mark.

The docs describe the text mark as invisible. It is applied while the model writes, by slightly shaping which tokens it picks, and adds no extra characters. SpaceXSI says the method uses SynthID Text, a technique published by Google DeepMind. A response header can indicate whether a given reply was watermarked, and detection is statistical.

The default was not always off. SpaceXSI had earlier told API customers that text watermarking would be on by default from November 7, citing compliance with the EU AI Act. The New Stack, citing an archive of the docs, reported that the written default appeared to change between November 9 and November 10, and that it was unclear whether the original default on policy ever fully took effect. The November 15 posts are the public confirmation, including from Musk, that the default is now off.

What to expect next

The practical test is whether API teams leave the mark off, turn it on, or split teams to handle countries with different rules, since the setting applies to every key on a team rather than to a single request. Watch whether SpaceXSI explains the reversal, whether EU rules force the mark back on for some traffic, and whether other labs keep default on watermarks. Audio watermarking, which the docs still describe as default on, is a separate setting to track.