A default that arrives on November 7
Most API changes ask developers to do something. This one asks them to do nothing, and that is the point. From November 7, every text response from seven Grok models carries an invisible watermark unless a team admin has already switched it off for that model. There is no parameter to send and no change to the request format. The one visible trace is a response header, x-watermark, that records whether that particular response was generated with the mark.
The scope is wider than prose. The documentation covers structured outputs and tool-call arguments as well as ordinary text, on the global endpoint and the regional ones, through both the Responses API and Chat Completions. A team that pipes Grok output into a database, a document store or another program's inputs is having all of it marked, not just the paragraphs a person eventually reads.
Marked by default. Checking a passage needs approval.
The asymmetry sits in the detection rules. Reading a header tells you whether one response you just received was watermarked. Testing a passage you already hold, perhaps one a contractor delivered or a model produced last month, is a different capability, and it is not public. SpaceXSI scores text against the watermark and returns one of three answers, watermarked, not watermarked, or uncertain, with longer text giving a stronger signal. Access to that detection API is granted by application.
The position for an ordinary reader is more limited than the headline suggests. An API customer can inspect the x-watermark header and establish whether a particular response was generated with the mark. But checking an arbitrary passage of text, perhaps one a contractor delivered or a model produced last month, requires the separate detection API, which SpaceXSI grants by application.
Even that detector does not return a definitive authorship verdict. Its three possible results are "watermarked", "not watermarked" and "uncertain". Longer text provides a stronger statistical signal, according to the documentation. A positive result is evidence of a watermark, not proof of who wrote the passage or how much human editing it received.
SpaceXSI publishes no false-positive rate, detection-rate benchmark or evaluation of how its watermark survives paraphrasing and editing. That leaves developers without quantitative evidence for judging the detector's reliability. OpenAI, by comparison, has published false-positive and detection-rate measurements for its own watermarking system, although those results cannot be transferred to xAI's implementation.
Opting out is a signature, not a setting
The off switch exists, and its design tells you what it is for. It lives in admin-only team settings, applies to every API key on the team at once, takes effect within five minutes, and is written to the audit log. Turning a model off requires the admin to confirm that an exception under applicable law applies to the team's use, and that the team is solely responsible for any marking, disclosure or other legal requirement. Some models may be locked on for a particular team. They appear in the settings dialog without an opt-out checkbox, meaning that team's administrator cannot disable their watermark through the Console.
Read plainly, this is not an ordinary preference toggle. A team that leaves the setting untouched gets the provider's default marking. A team that switches it off must have an administrator confirm that an applicable legal exception exists and accept responsibility for any remaining marking or disclosure requirements. That creates a meaningful compliance hurdle, even though the documentation does not explain why SpaceXSI chose this particular opt-out process.
The stated reason is one sentence in the documentation: SpaceXSI says it watermarks text output to comply with the EU AI Act. The legal timing provides context. Article 50(2) requires providers of generative AI systems to mark synthetic text and other generated content in a machine-readable and detectable format. Under the regulation's transitional provision, systems placed on the market before August 2, 2026 must take the necessary steps to comply by December 2, 2026. SpaceXSI's documentation does not name either provision or deadline. Its November 7 implementation is a company-stated compliance measure, not proof that the specific technical approach satisfies every applicable legal requirement.
The claim with no numbers attached
One sentence in the documentation deserves a slower read: watermarking does not change response quality or speed. Shaping which tokens a model picks is, by construction, a change to how text is generated. It may well be a change too small to measure in ordinary use, and other providers have made the same claim for the same technique. But the documentation offers no measurement, no evaluation set and no independent test behind the sentence. Until one appears, treat it as an assurance, not a result.
OpenAI's October 5 announcement takes a different approach for its API. API customers globally can opt in to text watermarking for selected models, but OpenAI says the API default will remain off. Its eligible ChatGPT and Codex text output in the EU is a separate rollout, not a global API default. The contrast is therefore specific but meaningful: xAI puts the obligation on the API provider by default, while OpenAI initially leaves the API decision with the customer. Both restrict access to their text-detection capabilities.
The broader industry is moving toward machine-readable signals for generated text, but providers are choosing different implementation paths. OpenAI has announced an opt-in default for API customers, while xAI is making watermarking the default for the seven listed Grok API models. Anthropic has also been associated with text-watermarking developments, although reported extensions to additional Claude models should not be treated as confirmed until supported by its own documentation. The meaningful comparison is not which company adopted watermarking first. It is who controls the default, who can inspect the signal, and who carries responsibility when it is switched off.
Sources
- [1] SpaceXSI (xAI) developer documentation, “Output Watermarking”, updated October 31, 2026 (primary source, read first-hand; seven models from November 7, x-watermark header semantics, detection by application, opt-out confirmation, locked-on models, no-change quality and speed claim)Read source
- [2] OpenAI, “Our approach to EU text provenance rules”, October 5, 2026 (read first-hand; API customers globally can opt in for select models and text watermarking remains off by default in the API; eligible ChatGPT and Codex text output in the EU is a separate rollout, not a global default; detector access limited to approved researchers and expert organizations; published detection measurements at a 1% target false-positive rate — about 80% of 200-token and 95% of 400-token passages)Read source
- [3] Regulation (EU) 2024/1689 (EU AI Act), Article 50(2), EUR-Lex: providers of AI systems generating synthetic audio, image, video or text content shall ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulatedRead source
- [4] Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex, inserting Article 111(4) into the AI Act: providers of AI systems generating synthetic audio, image, video or text content placed on the market before 2 August 2026 shall take the necessary steps to comply with Article 50(2) by 2 December 2026Read source