SpaceXSI, SpaceX’s AI unit, has made text watermarking on its Grok API opt-in. A SpaceXSI customer email, shown in a screenshot posted on X by @techdevnotes on 13 November, says: “As of November 9, 2026, your text outputs are watermarked only if a team admin turns it on”. Its current docs now open with: “Text watermarking is off by default”.
SpaceXSI’s earlier email said watermarking “is on by default for all teams and will be enabled on November 7, 2026”, as we reported on 31 October. From 7 November to 9 November is 2 days, by our arithmetic. Elon Muskwrote on X on 14 November: “Text watermarking with Grok is only on where required by law or requested by the user”.
Business Pill · The Model and the Harness
A short explainer on the difference between an AI model and the product built around it. It teaches the general idea only and says nothing about the watermark in this piece.
The key insight: As we read the two emails, SpaceXSI has moved the text-marking decision from its own default to its customers: on 31 October the watermark was on unless a team claimed a legal exception, and as of 9 November it is off unless a team admin turns it on. The switch came 23 days before the AI Act’s 2 December date for older systems, by our arithmetic.
The version of the watermarking page we read on 31 October said: “Starting November 7, 2026, responses from these models include a watermark. It’s on by default, and you can opt out per model”. It also said: “SpaceXSI watermarks text model output to comply with the EU AI Act”.
The page we read on 15 November says team admins “can turn it on per model in the SpaceXSI Console’s team settings”. The compliance line now reads: “Watermarking can help enable compliance with the EU AI Act”. Both versions list the same seven Grok text models, from grok-4.7 to three grok-4.20 variants.
Some terms stayed. Turning a model off still “asks you to confirm that an exception under applicable law applies to your use”, and the page still says “Some models may be locked on for your team”. The page we read does not say which models are locked on, or why.
The new page also names the method: “Our implementation uses SynthID Text, the text watermarking technique published by Google DeepMind in a Nature paper in 2024”. It adds an audio section: voice output from Text to Speech and Speech to Speech is watermarked by default with AudioSeal.
SpaceXSI’s Grok API text-watermark default: the 31 October email and docs set it on by default from 7 November; the 13 November email says text is watermarked only if a team admin turns it on as of 9 November. Article 111(4) of the EU AI Act, per the AI Act Explorer, gives older systems until 2 December 2026. Day counts are our arithmetic.
What Was Posted on X
Text watermarking is now off by default in the SpaceXSI API
The @techdevnotes account posted this at 22:33 UTC on 13 November with a screenshot of a SpaceXSI customer email titled ‘Text watermarking is now off by default in the SpaceXSI API’. The email says that as of November 9, 2026, text outputs are watermarked only if a team admin turns it on. The piece quotes the email from this screenshot and checks its terms against SpaceXSI’s public docs.
Text watermarking with Grok is only on where required by law or requested by the user
Elon Musk quoted that post at 16:39 UTC on 14 November, writing that text watermarking with Grok is only on where required by law or requested by the user. The post had about 1.6 million views when we read it. The piece sets these words against what SpaceXSI’s docs say for text and for audio.
Where the Law Requires a Mark
The 13 November email puts the legal call on customers: “We expect you will conduct appropriate reviews for your compliance with any laws. If you’re required to mark AI-generated text, turn watermarking on for the models you use”.
For audio, the docs set a location rule. Even for a voice model a team opted out, audio is watermarked when “The end user is in the EU, the EEA, or California”, and “The location comes from the IP address of the request”. The text sections of the page we read give no location rule; there, the switch sits with the team admin.
California’s AI Transparency Act asks a covered provider for a latent disclosure “in any image, video, or audio content, or content that is any combination thereof”. Text is not in that list. The chapter “shall become operative on August 2, 2026”, the code says.
The EU rule does name text. Article 50(2) of the AI Act, as shown in the AI Act Explorer, covers providers “generating synthetic audio, image, video or text content” and says outputs must be “marked in a machine-readable format and detectable as artificially generated or manipulated”, with exceptions such as “an assistive function for standard editing”.
Article 111(4), added by an amendment according to the Explorer, gives providers of systems “placed on the market before 2 August 2026” until “2 December 2026” to comply. That is 23 days after the 9 November switch, by our arithmetic.
Days between steps in SpaceXSI’s Grok API text-watermark policy, from its 31 October and 13 November emails and the AI Act Explorer’s Article 111(4) date. Day counts are our arithmetic.
How OpenAI, Anthropic and Google Handle It
OpenAIsaid on 5 October: “Text watermarking will remain off by default in the API”. For ChatGPT and Codex it is adding the watermark in the EU only, and said: “We are not making text watermarking a global default at launch”. Our OpenAI watermark piece covers its textGrain method.
Anthropictook the other route: “We’re applying watermarking globally at launch because we don’t yet have a durable way to scope it by region”. It says Claude’s text watermark “is a version of the SynthID-Text approach published by Google DeepMind”; our Anthropic watermark piece has the model list.
Google’s SynthID page says: “We’ve expanded SynthID to watermarking and identifying text generated by the Gemini app and web experience”. On text in the API, SpaceXSI’s default now matches OpenAI’s, and both differ from Anthropic’s global approach.
The Structural Read
As we read Musk’s post against the docs, “requested by the user” matches the per-model switch a team admin can turn on, and “where required by law” matches the audio rule for users in the EU, the EEA and California; for text, the docs tell customers to make that call themselves.
As we read the legal texts, the audio rule lines up with California’s law, which lists image, video and audio but not text, while the EU’s Article 50(2) names text as well; the docs we read do not apply the EU location rule to text.
SpaceXSI customer email, 13 November 2026 (screenshot posted on X by @techdevnotes)
“If you’re required to mark AI-generated text, turn watermarking on for the models you use.”
Three Implications
FOR DEVELOPERS ON THE GROK API Text is no longer watermarked unless a team admin turns it on per model, and the setting covers every API key on the team, the docs say.
FOR EU DEPLOYMENTS Article 50(2) names text, and Article 111(4) gives systems on the market before 2 August 2026 until 2 December 2026, per the AI Act Explorer.
FOR VOICE APPS Audio stays watermarked by default and, for end users in the EU, the EEA or California, stays watermarked even after an opt-out, the docs say.
The framework says: “The limiting factors are organizational readiness, governance, workflow redesign, training, and trust.”
As we read it, a default-on watermark was one way a model provider took on part of that governance work for its customers; the 9 November switch hands it back to them. The framework is the owner’s argument and does not discuss SpaceXSI’s watermark.
What Is Not Established
SpaceXSI gave no public reason for the change in the email or docs we read. The New Stack, which reported the switch on 10 November, wrote: “What remains unclear is why SpaceXSI abandoned its previously stated plans to make the feature automatic”.
Whether any Grok text was watermarked by default between 7 and 9 November, how many teams have turned it on, and which models are locked on are not established. The 13 November email is known from a screenshot; its terms match the public docs. SpaceXSI was not contacted for this piece.
This Week In The AI Supercycle
AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source
This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.
SpaceXSI’s Grok API text watermark was due on by default from 7 November and is opt-in as of 9 November, its emails say: 2 days, by our arithmetic. Audio stays watermarked by default and cannot be switched off for users in the EU, the EEA or California, while the AI Act’s text-marking date for older systems is 2 December 2026.
A note on sourcing. We compared SpaceXSI’s Output Watermarking developer page as we read it on 31 October 2026 with the page on 15 November 2026, and read the SpaceXSI customer emails shown in screenshots posted on X by @ai_for_success on 31 October and @techdevnotes on 13 November, plus Elon Musk’s 14 November post. The legal text comes from Articles 50 and 111 of the EU AI Act as shown in the AI Act Explorer (EUR-Lex was not fully available when we checked) and from the California AI Transparency Act on the state legislature’s site.
We also read OpenAI’s, Anthropic’s and Google’s own watermarking pages and The New Stack’s 10 November report. Day counts are our arithmetic. SpaceXSI was not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
