xAI Will Invisibly Watermark Text From Its Grok API Models by Default Starting Nov 7, Citing the EU AI Act
xAI announced today it is adding text watermarking to Grok. This appears to be rolling out globally. xAI was the last holdout, so the only model left that (as far as I know) does not watermark text is ChatGPT, and only ChatGPT in North America.
Context
Text from one more major model family becomes statistically traceable as machine generated, leaving few large providers that do not watermark text output.
Text watermarking nudges which tokens a model picks so that its output can later be recognized statistically, without adding visible marks. Andrew Curran noted that Anthropic, OpenAI, Google, Meta, Microsoft and Mistral signed the EU Code of Practice on Transparency of AI Generated Content, which commits them to watermark text, while xAI did not sign. The new xAI developer page, dated Oct 31, says the company now watermarks to comply with the EU AI Act.
Deep dive
xAI, which its developer docs now call SpaceXSI, published an Output Watermarking page on Oct 31. It says that starting Nov 7, 2026, text from Grok 4.7, Grok 4.6, Grok 4.5, Grok 4.3 and the three Grok 4.20 variants (reasoning, non reasoning and multi agent) carries an invisible watermark by default. The company gives compliance with the EU AI Act as the reason, and says the mark is applied on every API endpoint, including the global api.x.ai endpoint and regional ones.
According to the docs, the watermark is applied during generation by slightly shaping which tokens the model selects, so no characters, markup or hidden text are added. It covers ordinary text, structured outputs and tool call arguments in both the Responses API and the Chat Completions API. A new response header returns true when an output was marked. xAI says quality and speed are unchanged; it has published no accuracy or false positive figures.
Detection is statistical and returns watermarked, not watermarked or uncertain, with longer text giving a stronger signal. Access to the detection API is granted by application only. Team admins can opt out per model in the console, but must confirm that a legal exception applies to their use, and some models may be locked on.
Andrew Curran framed it as xAI being the last holdout and said ChatGPT in North America is now the only text he knows of that is not watermarked; that comparison is his own and was not checked here. TokenPost reports the change covers the API, not the consumer Grok apps.
What to expect next
Watch whether xAI extends the watermark to the consumer Grok apps on X and the web, and whether OpenAI follows for ChatGPT in North America. Also watch who gets access to the detection API, and whether independent researchers test how well the mark survives paraphrasing, translation and short outputs, the usual weak points of statistical text watermarks. The Nov 7 start date and the opt out rules are the next concrete checkpoints for developers on the API.