xAI turns on default text watermarking for the Grok API

xAI turned on invisible text watermarking by default across its Grok API on November 7, 2026. Seven current Grok model variants now ship watermarked output unless a team admin opts out per model, which makes this the first confirmed default-on watermark policy at a major frontier API, landing three weeks before the EU AI Act's December 2 marking deadline.

Key facts

• Starting November 7, 2026, text responses from 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 include a watermark by default, per xAI's own developer documentation.
• Watermarking applies on every xAI API endpoint, including api.x.ai and regional endpoints, across both the Responses API and Chat Completions API, and covers ordinary text, structured outputs, and tool-call arguments.
• Integrators can check a single response by reading the x-watermark response header: a value of true means that response was generated with the watermark.
• Detection is statistical and gated: xAI scores text and reports watermarked, not watermarked, or uncertain, and access to the detection API is granted by application form only.
• The opt-out is team-admin-only, per model, in the xAI Console: it applies to every API key on the team, propagates within 5 minutes, is audit-logged, and requires confirming that an exception under applicable law covers the customer's use.
• xAI states the watermark exists to comply with the EU AI Act and separately asserts it does not change response quality or speed. The compliance rationale is in the doc; the quality claim is the company's own assurance, not an independently verified property.

What happened

The policy is documented on a page titled Output Watermarking in xAI's developer docs, and the effective date is stated in the document itself: "Starting November 7, 2026, responses from these models include a watermark. It's on by default, and you can opt out per model." The list covers seven model variants: 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. Coverage spans text generation on both the Responses API and the Chat Completions API, including structured outputs and tool-call arguments, on every endpoint including api.x.ai and the regional endpoints.

The mechanics matter as much as the date. The watermark is applied during generation: the doc says it "slightly shapes which tokens the model picks," and that it adds no characters, markup, or hidden text to the response. There is no request parameter, and the response format is unchanged except for the x-watermark header. Detection is a separate, statistical process: xAI scores a piece of text against the watermark and returns watermarked, not watermarked, or uncertain, with longer text producing a stronger signal.

Access to that detection capability is not open. xAI grants access to the detection API by application form, so a builder who suspects an output came from a watermarked model cannot verify it independently today. What any integrator can do is read the per-request header on their own traffic, which tells them whether xAI generated that response with the watermark on.

The opt-out structure is the part most likely to surprise platform teams. It is not a developer-level or per-request setting. A team admin flips it per model in the Console, it applies to every API key on the team, changes propagate within 5 minutes, and every toggle lands in the team's audit log. Confirming the opt-out requires asserting that an exception under applicable law applies to the customer's use and accepting sole responsibility for any marking or disclosure duties. The doc also notes that some models may be locked on for a team, with no Console path to turn them off.

Two details from the doc deserve neutral treatment. First, the page repeatedly names the operating entity "SpaceXSI," which is unusual branding inside xAI's own documentation; the policy text, not the entity name, is the news here. Second, the FAQ answers "Does watermarking affect quality or speed?" with a flat no. That assurance sits in tension with the same page's statement that the watermark shapes token selection. Both statements come from xAI. The quality claim is unverified until independent evals compare watermarked and unwatermarked outputs from the same model, and DeAI treats it as a claim, not a fact.

Why it matters

For any builder calling the Grok API, directly or through an aggregator, the default changed underneath them on November 7. Outputs are now statistically marked whether or not the downstream product asked for provenance. Three practical consequences follow. Evaluation pipelines that compare cached outputs or expect deterministic regeneration may see token-level shifts across the November 7 boundary. Products with their own disclosure obligations now have a per-request signal, the x-watermark header, that they can surface or log. And teams that want the watermark off cannot get there with an API key alone: the control belongs to team admins, requires a legal-exception confirmation, and may be unavailable entirely for locked models.

The timing is the tell. The EU AI Act's Article 50(2) marking duty becomes applicable on December 2, 2026 for generative systems placed on the market before August 2, 2026. xAI's November 7 default-on, with the EU AI Act named as the reason in its own docs, is the first confirmed provider response to that runway. DeAI's week-ahead brief for November 2-8 flagged exactly this as the watch item, including an unverified claim that Anthropic is expanding text watermarks on additional Claude models. If xAI moved first under Article 50 pressure, the remaining frontier providers are the obvious next candidates, and the December 2 date is 25 days out.

Background

The EU AI Act's transparency duties took effect on August 2, 2026, when the EU AI Office also gained investigative powers and fining authority up to EUR 15 million or 3% of worldwide turnover. DeAI covered that enforcement phase in September, including the machine-readable marking duty and the December 2 grace date for systems already on the market. November 7 is the first confirmed default-on response from a major lab inside that window.

The change also lands on an xAI policy surface builders already track warily. The company's consumer data-training defaults have churned repeatedly since launch, which is why every xAI policy claim on this publication is dated and re-verified. And prior coverage has noted that watermarking techniques can shift model refusal behavior, a research finding that gains relevance when the mechanism is applied at generation time, as xAI's doc describes, rather than appended to finished text.

What's next

Three things to watch. Whether other frontier APIs announce equivalent default-on watermark policies before December 2, 2026, and whether they copy xAI's team-admin opt-out structure or offer anything finer-grained. Whether independent researchers publish evals comparing watermarked and unwatermarked outputs of the same Grok model, which is the only path to settling the quality assertion. And whether the list of locked-on models grows, since a lock removes the only off switch xAI currently documents. If your evaluation suite compares Grok outputs across the November 7 boundary, that comparison now spans two generation regimes.