Grok watermark
Grok puts a visible mark on generated images and video and documents nothing for text. Its maker is also the one major provider that did not sign the EU transparency code of practice.
What the evidence says
No text watermark is documented for Grok. The provider’s FAQ raises watermarking only under image and video generation, the API documentation for those endpoints mentions neither watermarking nor provenance, and no detection tool is published for any modality. Nor is this a denial: asked about its plans in August 2026, the company did not respond. This entry records silence rather than a statement.
A naming note, because it dates quickly otherwise. The company behind Grok is now SpaceXAI, following SpaceX’s acquisition of xAI during 2026; its policies and documentation carry the new name while the product keeps the name Grok. Material published before the rename appears under xAI.
Images and video are marked, but visibly. The FAQ states that generated images and videos include a Grok watermark to indicate the content was created with AI, and that there is no setting to remove it. A visible label and a cryptographic watermark are different things: the first is seen by the viewer and croppable by an editor, the second is embedded in the content. Only the first is documented, no C2PA membership is listed for the company, and neither the mark’s appearance nor the date it was introduced is described in any primary source, so this page does not describe them.
The acceptable use policy forbids stripping, altering or circumventing embedded provenance metadata or watermarks. Read it for what it is: a rule imposed on users, filed among the things that harm the service, alongside scraping and rate-limit evasion. It is not a statement that the models embed provenance signals, and it says nothing about text.
The regulatory position is what makes this entry distinctive. SpaceXAI did not sign the European Commission’s code of practice on transparency of AI-generated content, unlike Anthropic, Google, Microsoft and OpenAI. That is not an infringement in itself — the Commission is explicit that signing is voluntary and that not signing does not constitute non-compliance — but a non-signatory has to demonstrate compliance by other means. Article 50(2) of the AI Act names synthetic text alongside audio, image and video, has applied since August 2026, and gives systems already on the market until 2 December 2026 to meet the marking obligation.
A common misreading
The acceptable use policy’s ban on stripping “embedded provenance metadata or watermarks” is often read as proof that Grok hides signals in everything it produces, text included. It is a rule for users, not a description of the output, and the only marking the company documents is a visible label on images and video.
Sources
- Grok FAQ: image and video generationSpaceXAI · · Provider documentation
- SpaceXAI Acceptable Use PolicySpaceXAI · · Provider documentation
- Strong backing for the Code of Practice on Transparency of AI-generated ContentEuropean Commission · · Independent report
- Signing the Code of Practice on Transparency of AI-generated ContentEuropean Commission · · Independent report
Frequently asked questions
Does Grok watermark its text?
Nothing is documented. The FAQ covers images and video only, no detection tool is published, and the company has not said either way when asked. That is silence rather than a denial.
What about Grok images and video?
They carry a visible Grok mark, and the FAQ says there is no setting to remove it. That is a label rather than an embedded cryptographic watermark, and removing it is prohibited by the acceptable use policy.
Does the EU AI Act require Grok to mark text?
Article 50(2) names synthetic text and has applied since August 2026, with a deadline of 2 December 2026 for systems already on the market. Not signing the voluntary code of practice is not, by itself, non-compliance.
Related tools
Working on your own draft?
GenPolish rewrites text you already have, showing every change before you accept it. It does not add or remove provenance signals, and it makes no claim about detectors.