If your app generates text through the Grok API, that text is now quietly labelled as machine-made. On 7 November 2026, the change went live: text generated by seven Grok models on the SpaceXSI API is now watermarked by default, with the markup invisible to human readers but readable to the systems that check for it. If you use these models in your own product, you need to know how to check whether your calls are watermarked, and how to opt out if your own compliance setup demands it.
What changed
- Which models: grok-4.7, grok-4.6, grok-4.5, grok-4.3 and three grok-4.20 variants — seven in total.
- Where it applies: the API only. This change does not apply to the Grok app or the Grok website, and it does not change image generation.
- How it is signalled: a new response header,
x-watermark: true, tells you whether a response was generated under the watermark. - What the watermark is: a subtle, token-level marker — embedded into the text itself, not an addition at the end, and intended to survive normal pasting, forwarding and reformatting.
Why now: the EU AI Act deadline
The driver is Article 50(2) of the EU AI Act, which requires providers of general-purpose AI models to label their output as AI-generated so users and others know what they are reading. The requirement applies gradually (an August 2026 start date for full applicability of general-purpose AI obligations), and deployers have been given a transition period, and they must comply by 2 December 2026. Switching the watermark on by default in November puts Grok's API inside that deadline with a few weeks to spare. Reporting on the change is covered by FourWeekMBA and Maccus.
Step by step: check your own API traffic
- Make a normal API call to any of the seven models listed above, exactly as your app does today — no new parameters are needed.
- Inspect the response headers (not the body) in your HTTP client or logs, and look for
x-watermark: true. If the header is present and true, that response was generated with the watermark on. - Log it per model. Because the setting is controlled per model, check each model your app uses separately — do not assume they all behave identically.
- If you need the watermark off: the opt-out lives with your team admin, who can turn watermarking off per model. Be aware of the condition attached: opting out means accepting sole responsibility for compliance — the legal duty does not disappear, it moves to you.
- Re-check after any model change. If you switch models or add a new one, repeat step 2 — a new model may come with watermarking on by default even if your existing ones were opted out.
Should you opt out?
For most developers, the honest answer is no — leave it on. The watermark is invisible to readers, it is exactly what EU law asks for, and keeping it costs you nothing while removing compliance risk. The opt-out exists for teams with their own marking or disclosure systems already in place — and even then, the "sole responsibility" condition means your legal position should be checked before anyone flips that switch. The deadline to have an answer, one way or the other, is 2 December 2026.
Sources: FourWeekMBA, "SpaceX AI will watermark Grok text by default for EU AI Act", November 2026; Maccus, "Grok text watermark", November 2026; EU AI Act Article 50(2) transition deadline, 2 December 2026. Hero image: hands typing on a laptop keyboard, Wikimedia Commons.
