Takeaway: Starting November 7, 2026, SpaceXSI (the docs brand formerly associated with xAI) turns invisible text watermarking on by default for seven Grok models on every SpaceXSI API endpoint — including https://api.x.ai and regional endpoints — to comply with the EU AI Act. There is no request parameter and no required code change. The response header x-watermark: true means that reply was generated with the watermark. Team admins can opt out per model in the SpaceXSI Console after confirming a legal exception and accepting sole responsibility. This article covers API scope only; SpaceXSI’s docs do not clearly say the consumer Grok app or X posts are watermarked.
Primary source: SpaceXSI developer docs — Output Watermarking. Broader EU transparency context: Article 50 AI Act FAQs. Independent write-ups such as The Tech Signal summarize the same November 7 default without adding claims beyond the docs.
What turns on November 7, 2026
Watermarking applies to text output from text-generation requests for each model that has watermarking turned on for your team. From November 7, 2026, responses from these models include a watermark by default (you can still opt out per model where the console allows it):
grok-4.7grok-4.6grok-4.5grok-4.3grok-4.20-reasoninggrok-4.20-non-reasoninggrok-4.20-multi-agent
Coverage is broad on the text side: the watermark applies to outputs from every text-generation API, including the Responses API and the Chat Completions API. It covers ordinary text responses, structured outputs, and tool call arguments — anything in the generated text you receive back.
How the invisible watermark works
The watermark is not visible markup. SpaceXSI says it is applied while the model generates text: it slightly shapes which tokens the model picks, so it adds no characters, markup, or hidden text to the response. Detection is statistical: SpaceXSI scores a piece of text and reports whether it is watermarked, not watermarked, or uncertain. Longer text gives a stronger signal.
Access to watermark detection is granted by application; the docs point teams to a request form for the detection API. Separately, for any single API call you can check whether that response was watermarked by reading the x-watermark response header on Responses or Chat Completions. x-watermark: true means watermarked; any other value, or no header, means it was not.
SpaceXSI states watermarking does not change response quality or speed. There is still no request parameter — integrations keep working the same way with or without the mark.
Confirmed vs out of scope
| Detail | Status (per SpaceXSI docs) | Notes |
|---|---|---|
| Default on from Nov 7, 2026 for seven listed models | Confirmed | On by default; opt-out available where unlocked |
| Applies on every SpaceXSI API endpoint (global + regional) | Confirmed | Includes https://api.x.ai |
| Responses API + Chat Completions text / structured / tool args | Confirmed | Generated text surfaces listed in docs |
| No request parameter / no required code change | Confirmed | x-watermark header is the per-response signal |
| Invisible token-shaping watermark; no added characters | Confirmed | Statistical detection; longer text stronger |
| Team-admin Console opt-out; ~5 minutes; audit logged | Confirmed | Confirm legal exception + sole responsibility |
| Some models may be locked on | Confirmed | Locked models show without a checkbox |
| Consumer Grok app / X posts watermarked | Not clearly stated | Treat as API-scope only until docs say otherwise |
| Detection API open to all developers | No | Access by application |
Opting out in the SpaceXSI Console
Watermarking is controlled per model for your whole team. Only team admins see the setting. The high-level path in the docs:
- Open Team settings and find the Model output watermarking row (On / Off / Custom).
- Open the settings icon, then clear the checkbox for each model you want to opt out (or use select-all).
- Save. If you turned a model off, confirm with Opt Out.
- Changes apply to new requests within about 5 minutes. Opting back in starts watermarking again on new requests.
Opting out requires confirming that an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure, or other legal requirement. The setting applies to every API key on the team. Every opt-in and opt-out is recorded in the team audit log.
Some models may be locked on for your team — they appear in the dialog without a checkbox. Questions about locked models go to your team admin, then support@x.ai if needed.
Practical checklist for API teams before November 7
- Inventory models: Map which of the seven default-on models your production and staging keys actually call.
- Decide default posture: Leaving watermarking on is the path of least friction for EU AI Act transparency obligations; opt-out is a legal+ops decision, not a performance tweak.
- If you may opt out: Have counsel confirm whether a legal exception applies before an admin clicks Opt Out; document who approved it.
- Wire header checks: Log or assert
x-watermarkon Responses/Chat Completions in staging so you can prove whether a reply was marked. - Tool-calling apps: Remember tool call arguments are in scope — treat them like any other generated text for disclosure policies.
- Detection access: If you need third-party verification (compliance, trust & safety, research), apply for the detection API early; it is not open by default.
- Do not over-claim product scope: Update internal runbooks for API traffic only unless SpaceXSI later documents consumer Grok / X surfaces.
- Audit readiness: Know where Console audit logs live so you can show when watermarking was toggled.
Why publishers and SEO teams should care
Even if you never call the Grok API yourself, provenance rules shape what AI intermediaries can assert about generated copy — and how platforms may later label or filter synthetic text. That sits next to crawler and referral hygiene we already cover: robots.txt choices for AI crawlers, tracking AI referral traffic in GA4, and how search-like ChatGPT volume differs from clicks in our ChatGPT search ≈20% of Google query volume note.
Watermarking is text provenance: a statistical signal that content came from a watermarked model run. It is not the same as proving which model weights answered. For that adjacent cluster, see our Anthropic FlashProve coverage of model-identity / zkSNARK-style inference proofs — useful contrast when stakeholders mix “was this AI text?” with “which model produced it?”
What this is not
- Not a claim that consumer Grok chat or posts on X are watermarked — SpaceXSI’s page documents API text generation.
- Not a visible steganography layer you can spot in the string (no zero-width characters described).
- Not a quality or latency tax according to SpaceXSI’s FAQ.
- Not automatic access to SpaceXSI’s detector for every customer.
FAQ
Do I need to change my API code before November 7?
No. SpaceXSI says watermarking has no request parameter, and the response format is unchanged apart from the x-watermark header. Your existing Responses or Chat Completions integration keeps working.
Which Grok models get the default watermark on November 7, 2026?
According to SpaceXSI docs: grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning, and grok-4.20-multi-agent.
How do I tell if a single response was watermarked?
Read the x-watermark header on the Responses API or Chat Completions API response. true means that output was generated with the watermark; any other value or a missing header means it was not.
Can my team turn watermarking off?
Yes, if you are a team admin and the model is not locked on. Use Team settings → Model output watermarking, opt out per model, confirm the legal exception and sole responsibility, and expect changes on new requests within about five minutes. The choice applies to every API key on the team and is audit-logged.
Does this watermark the consumer Grok app?
SpaceXSI’s Output Watermarking docs describe API text generation endpoints. They do not clearly state that the consumer Grok app or X posts are watermarked, so treat consumer surfaces as out of scope until official docs say otherwise.