AI writing watermarking

Anthropic has announced it will begin "watermarking" LLM-produced text to comply with the EU AI Act, which requires the labeling of content developed by artificial intelligence. While the company published an explainer, I asked AI to demonstrate such watermarking by italicizing elements of the watermark in its response below.

As you can see, the watermarks are most commonly modifiers, and then nouns and verbs, that could be substituted or removed without changing the meaning. Such facts as 30 and 28 percent can't be substituted and therefore won't be watermarked.

An LLM doesn't compute one "correct" next word. At each step it produces a probability distribution — maybe "quick" at 30 percent, "rapid" at 28 percent — and samples from it. At most positions in fluent prose, several continuations are effectively tied in quality. That slack is where the watermark lives.

A pseudorandom function — seeded by a secret key plus the preceding few tokens — quietly influences which of the near-tied candidates gets picked. Nothing is added to the text; the watermark is the pattern of choices. Over hundreds of tokens, a detector holding the key can check whether the text aligns with the keyed pattern far more often than chance. Any single word proves nothing; the statistics across the whole passage do.

Where prose is genuinely flexible, the watermark hides in choices a careful writer would consider interchangeable. Where prose isn't — facts, quotes, technical terms — the watermark simply can't exist.