Anthropic Uses Invisible Text Watermarks in Claude for EU Rules

Anthropic has added invisible text watermarks to Claude outputs from models released since Aug. 2 to meet EU AI Act transparency rules, Politico reports.

The watermark is built into the text itself. Anthropic says it is not visible to readers, does not change the meaning or readability of responses, and remains with text when it is copied and pasted.

The method works by changing some word choices in places where the company says the choice does not affect meaning. Anthropic says the watermark is used less in factual passages and code, where word or symbol changes could affect accuracy.

The system is meant to show that Claude was likely involved in generating a piece of text. Anthropic says light editing may leave the signal in place, while a full rewrite can remove it.