xAI to switch on invisible watermarks by default in Grok text APIs from November 7, citing the AI Act
xAI's official API documentation, published under the name SpaceXSI, announces that from November 7, 2026 all text produced by seven Grok models will carry an invisible watermark by default. The mechanism adds no characters or markup to the content: it slightly nudges token selection during generation, which allows statistical detection with three possible outcomes: watermarked, not watermarked or uncertain. The `x-watermark` response header (value `true`) tells the caller whether the output was marked, while the detection API is available on request by filling in a form.
It applies to every text generation API, including Responses and Chat Completions, and also covers structured outputs and tool-call arguments. According to TokenPost (October 31), Grok's consumer products are not affected. The documentation lists 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. TokenPost counts five because it groups the three grok-4.20 variants into a single entry (the page doesn't explain how the SpaceXSI name relates to xAI). Team admins can turn the watermark off model by model from the Console; the change applies to all of the team's API keys within five minutes, but some models have the watermark locked and it cannot be removed. Every enable and disable action is recorded in the audit log.
The reason given in the documentation is compliance with the European AI Act. Article 50 of the regulation requires providers of generative AI systems to mark content in a machine-readable, detectable format. The documentation states that the watermark affects neither the quality nor the speed of responses, but no independent testing backs this claim. Nor is there any documentation on how robust the signal is against paraphrasing, translation or editing, or on the false-positive rate. It's unclear which models have the watermark locked and who will get access to detection. The sources also disagree on when it can be disabled: TokenPost mentions legal and compliance exceptions, while the documentation sets no conditions.
In short, xAI has announced a technical answer to the AI Act, but the lack of external evaluation and the ambiguities around how the watermark is managed call for caution from developers who rely on these APIs.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read xAI's official documentation (docs.x.ai) and confirmed the start date, models, covered APIs, mechanism, x-watermark header, on-request detection, per-model opt-out with audit log and the AI Act rationale, with verbatim quotes. Independent confirmation: TokenPost, 31/10/2026 (date, default setting, models, Console opt-out, consumer products excluded). Discrepancies between the two sources are flagged. I discarded a post on X linking the news to an alleged Claude watermark: unverified.
- Incertezze
- No independent check of xAI's claim that the watermark doesn't affect quality or speed. Robustness to paraphrasing, translation or editing and the false-positive rate are undocumented. The sources disagree on the number of models (seven vs five) and on the conditions for disabling it. It's unknown which models are locked and who will get detection access. The link between SpaceXSI and xAI isn't explained. The exact announcement date isn't stated: the confirming source is from October 31.
- Perché pubblicarla
- It's one of the first cases of a major model provider enabling a statistical text watermark by default via API, explicitly tied to the AI Act. It directly affects European and Italian developers and companies building on Grok, and raises concrete questions about detectability, content control and transparency, topics close to this site, which already shows the AI Act badge on every article.