xAI updated its API documentation on October 31 to announce that, starting November 7, text generated through the Grok API will carry an invisible statistical watermark by default, with no code changes required from developers. A claim circulating on social platforms over the past few days says Grok text watermarking is being “rolled out globally.” According to the documentation and several outlets relaying it, this round covers the API only; the Grok products aimed at ordinary users are unaffected.
The scope is five versions: grok-4.7, grok-4.6, grok-4.5, grok-4.3 and grok-4.20. grok-4.20 comes in reasoning, non-reasoning and multi-agent variants, which is why some outlets count seven models. A price list checked by a third party also includes grok-build-0.1, which does not appear on the watermark list; the documentation does not say why.
The watermark sits in word choice
The approach belongs to the same family as the text watermarks used by OpenAI and Anthropic. During generation it slightly nudges which tokens are selected. No single sentence looks unusual, and a signal can be detected statistically only once the text is long enough. There is no visible label and no hidden characters, so copied text looks like any other text.
Coverage is broader than chat replies. Ordinary responses, structured output and the arguments the model generates for tool calls all count, across both the Responses API and the Chat Completions API. Developers can check the x-watermark response header; a value of true means that output was watermarked.
To test an arbitrary piece of text, you have to apply separately for access to a detection tool, which returns one of three results: watermarked, not watermarked, or uncertain. Short text such as a one-line reply or a small JSON payload tends to land in “uncertain.” xAI says the watermark does not affect length, quality or speed. It has not disclosed accuracy, false-positive rates, or whether the mark can still be detected after paraphrasing or translation.
Turning it off comes with conditions
The setting lives under team settings in the console. Only admins can change it, it is toggled per model, and it applies to every API key on the team; individual keys cannot be configured on their own. Some models may be locked, in which case you have to contact the support team.
The more practical constraint is legal responsibility. To switch the watermark off, a team must confirm that a legal exception applies, and the team itself then carries any remaining labeling and disclosure obligations. Every change is written to the audit log. xAI cites the transparency requirements of the EU AI Act, whose relevant rules began to apply on August 2. Systems already on the market before that have a transition period until December under Article 50(2) as amended, and violations can draw fines of up to 15 million euros or 3% of global annual turnover.
The three companies each use different defaults
Set side by side, the differences come down to the default setting and the scope of coverage:
| Company | Consumer product | API |
|---|---|---|
| Anthropic | Watermarked worldwide | Watermarked |
| OpenAI | Eligible users in the EU | Can be enabled manually worldwide; off by default |
| xAI | Not covered in this round | On by default; can be turned off after declaring a legal exception |
When OpenAI announced textGrain in October, it published detection figures: at a 1% false-positive rate, a 400-token passage was detected about 95% of the time, dropping to 17% after a quarter of the words were replaced. For Grok’s watermark, there is no comparable public number on how well it survives rewriting.
The claim that “xAI is the last one, and only ChatGPT in North America is left without a watermark” comes from social media posts, and some outlets repeated it. Judged by product scope it does not hold up cleanly: consumer Grok is not included in this round, and as of October Microsoft Copilot has no public documentation describing a text watermark.
Questions remain about xAI’s change of stance as well. Earlier this year, media reported that it had not signed the EU’s code of practice on transparency for AI-generated content. The new documentation does not say whether it has signed, nor whether the watermark follows a common standard that third parties can read. Developers integrating the Grok API may want to check before November 7 where their output ends up; schools, publishing and content moderation are the settings most likely to run into detection results.
Sources: xAI API documentation, Crypto Briefing, TokenPost, CocoLoop, Data Studios; documentation details were cross-checked against media accounts, and the Act’s provisions should be read against the European Commission’s public explanations.