Definition
LLM text watermarking hides a signal in AI-generated text by slightly biasing which tokens the model picks while it writes. A reader sees normal prose. A detector that knows the scheme can count how often the favoured tokens appear and estimate whether the text came from a watermarked model. Nothing is attached to the text: no metadata, no hidden characters. The watermark is the word choice itself.
How Detection Works
Detection is statistical. A detector returns something like "watermarked", "not watermarked" or "uncertain", and its confidence grows with length. Short replies, heavy human editing, translation and paraphrasing all weaken the signal. A watermark is evidence of origin, not proof, and only the company that holds the key can run a reliable check.
Example: The Grok API Watermark
xAI, which its developer docs now call SpaceXSI, says it will watermark text from its API by default from November 7, 2026, "to comply with the EU AI Act". The docs list grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the grok-4.20 reasoning, non-reasoning and multi-agent models. It covers plain text, structured outputs and tool-call arguments. Each response carries an x-watermark header, and x-watermark: true means the output was watermarked. Team admins can opt out per model, but only after confirming a legal exception applies. Access to the detection service is by application. Details are in the Grok API watermark FAQ.
The docs do not cite an article number. Press coverage connects it to Article 50 of the AI Act, which deals with marking AI-generated content. That link is the press's reading, not xAI's.
Commonly Confused With
- Content credentials (C2PA): signed metadata attached to a file, mostly images and video. Easy to strip, but verifiable when present.
- AI text detectors: classifiers that guess whether any text "looks" AI-written. They need no key and are far less reliable.
- Image watermarks such as Google's SynthID for images work on pixels, not word choice.
Why It Matters for Brands
If your product descriptions, support replies or marketing copy are generated through a watermarking API, they carry the mark. That is usually fine. It matters if you have disclosure obligations, or if a platform starts down-ranking or labelling detected AI text. Know which of your vendors watermark by default. For adoption across vendors, see AI content watermarking adoption.