Grok API Text Gets an Invisible Watermark on November 7: What Changes and How to Opt Out

Text from xAI's Grok API will carry an invisible watermark starting November 7, 2026. The company's developer docs, which now brand the platform as SpaceXSI, say the mark is on by default for seven Grok models, needs no code change, and can be turned off per model by a team admin in the console. xAI says the reason is the EU AI Act.

This matters because xAI had been described as the holdout. Earlier in 2026 Anthropic announced invisible watermarks for Claude text, and in October we covered OpenAI's textGrain plan for EU ChatGPT and Codex. With Grok joining, the big three US text-model vendors are all moving the same way. Our earlier explainer asked whether all AI models will watermark output; this is a large step toward yes.

TL;DR: what the docs say

Which Grok models get the watermark?

The docs list seven models: 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. It applies to every text generation API, including the Responses API and the Chat Completions API. For background on the newest model, see our post on Grok 4.7 and its reported 2.1T parameters.

The docs frame the watermark as something that "lets SpaceXSI later recognize text as model-generated." The docs say the mark is applied while the model generates text, slightly shaping which tokens it picks, so it adds no characters or hidden markup. Detection is statistical: SpaceXSI scores a text as watermarked, not watermarked or uncertain, and longer text gives a stronger signal. Access to a detection API is granted by application through a form linked from the docs.

How does the x-watermark header work?

Each Responses API or Chat Completions API response can be inspected for an x-watermark header. true means the output for that request was generated with the watermark. Any other value, or no header at all, means it was not.

That makes the behavior testable in your own pipeline. A simple check after November 7:

curl -i https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"grok-4.7","input":"Write a haiku about the ocean."}' | grep -i x-watermark

We did not run this call ourselves, because the feature starts after publication. The header name and meaning come directly from the docs, so confirm against your own responses once the date passes.

How do admins opt out?

Watermarking is controlled per model for the whole team. Team admins can opt out in the SpaceXSI Console, according to the docs. That means a developer on a team cannot switch it off for a single key, and an opt-out on one model does not cover the others. Check each model you use, especially if you mix grok-4.7 with older grok-4.20 variants. Changes apply to new requests within about five minutes, every opt-in and opt-out is recorded in the team audit log, and some models may be locked on so they cannot be turned off from the Console. Opting out also asks you to confirm that a legal exception applies to your use and that you are solely responsible for any marking or disclosure the law requires.

Should you opt out? The reasons to stay opted in are regulatory comfort and provenance. If you serve EU users or publish generated text, a watermark can be evidence that you labeled machine output. The reasons to opt out are narrower: you want output free of any statistical signal, or you worry about downstream tools that might flag your text. Our explainer on how text watermarking works covers what these marks do and do not survive, and the developer impact guide covers integration questions.

Why is xAI doing this now?

The docs state the reason plainly: compliance with the EU AI Act. Article 50 of the Act sets transparency duties for providers of systems that generate synthetic text, requiring outputs to be marked in a machine-readable format and detectable as artificially generated, where technically feasible. Obligations are phasing in, and providers have been picking technical approaches.

The docs apply the watermark on every endpoint, including the global one, not only EU-region traffic. That differs from approaches where EU users get the mark and others do not. Anthropic's and OpenAI's own postures differ in detail, so compare them in our posts linked above rather than assuming they match.

What was reported before, and what changed?

Our research turned up conflicting earlier coverage. Press and commentary in August described xAI as not having signed the EU code of practice and as unwilling to embed text fingerprints. The developer docs now describe a default-on watermark. The docs page is marked "Last updated: October 31, 2026," and posts on X pointing at it, some calling xAI the last holdout among major labs, started circulating around November 1 and 2. Treat the August reports as superseded by the vendor's own documentation, which is the primary source.

Still unconfirmed, and worth watching:

• Detection access. A detection API exists but is granted by application. Whether regulators, publishers or customers will routinely get access is unknown.
• Robustness. The docs do not give figures on how well the mark survives paraphrasing or editing. No vendor has published strong numbers; see our look at removal and the law.
• Chat products. The docs cover the API. Whether Grok on X or the apps get the same mark is not stated in the page we read.

How does this compare with Claude and ChatGPT?

The three vendors are converging on the same goal, an invisible text mark tied to EU rules, but the packaging differs. Use this table as a map to our detailed posts, not as a spec.

The notable difference on the xAI side is scope. The docs say the watermark is enabled on every endpoint, global and regional, and that it is on by default. Whether another vendor limits its mark to EU traffic or to opt-in use is covered in the posts above, so check them rather than assuming symmetry.

What are people asking about it?

Will watermarked text read differently? The docs say integrations work the same way with or without the mark. The docs also say watermarking does not change response quality or speed, though they publish no measurements behind that claim.

Can the mark be removed? Research and commentary generally show text marks weaken under heavy paraphrasing, translation or manual rewriting. xAI's docs make no robustness claim. Our guide on watermark removal, right and wrong walks through what is legal and what is not, and the EU AI Act and DMCA explainer covers the legal side.

Does it affect tool calls and structured output? Yes, per the docs: the mark covers ordinary text, structured outputs and tool call arguments. If your pipeline compares generated JSON byte for byte, for example in caching or deduplication, test it after the switch. Do not assume a schema-valid JSON string is unchanged.

Who can detect it? SpaceXSI says the mark lets it later recognize text as model-generated, and the docs say outside access to the detection API is by application.

Is this a safety feature? No. It labels origin. It does not stop a model from producing harmful output or an agent from taking a dangerous action.

Context: why watermarking spread this year

Text watermarking moved from research into product policy in 2026 for one main reason: regulation set a date. Article 50 of the EU AI Act asks providers to mark synthetic output in a machine-readable way where feasible, and large vendors found it simpler to ship one global behavior than to branch by region. That logic explains why xAI's docs apply the mark on the global endpoint as well as regional ones.

The second driver is trust. Publishers, schools and platforms keep asking whether a passage was machine written. Provenance marks give vendors a way to answer for their own models, even though they cannot identify text from other sources. Our posts on marketers and on students and teachers cover how this lands outside engineering teams.

What should developers do before November 7?

1. Decide the policy per model. List which of the seven models your team uses and whether opt-out is appropriate for each.
2. Log the header. Capture x-watermark in request logs so you can later prove which outputs were marked.
3. Tell downstream users. If you pass Grok text into publishing, education or customer-facing flows, note that the text may be machine-recognizable.
4. Avoid text-cleaning assumptions. Because the mark lives in token choice rather than added characters, simple whitespace or Unicode cleanup should not "remove" it, but heavy rewriting may weaken it. Test, do not assume.
5. Review contracts. If a customer requires unmarked output, opt out and document the decision.

If you run agents that act on model output, remember watermarking is a labeling tool, not a safety control. For agent guardrails, AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

Bottom line

The vendor's own docs confirm the headline: Grok API text gets an invisible, default-on watermark from November 7, 2026, with a per-model opt-out and an x-watermark header for checking. What remains open is how detectable and durable the mark is, and who can read it. For builders, the action is small: audit your models, capture the header, and make a deliberate opt-in or opt-out call before the switch flips.

Related reading

• Anthropic's invisible watermarks for Claude
• OpenAI textGrain for EU ChatGPT and Codex
• Will all AI models watermark output?
• How AI text watermarking works
• Is removing an AI watermark illegal?
• AI watermarking: impact on developers
• Grok 4.7 and the SpaceX training data

Details are accurate as of November 2, 2026, from the SpaceXSI output watermarking docs. The page was last updated October 31, 2026; check it for changes.