Anthropic adds invisible watermarks to Claude's text output

Anthropic is rolling out invisible watermarking for Claude-generated text, based on Google DeepMind's open-source SynthID-Text approach, which embeds detectable patterns via word-choice probabilities. It's paired with C2PA marking for Claude-processed images, introduced to comply with the EU AI Act's requirement that synthetic content carry machine-readable marks. Critics question whether the method can avoid subtly altering word choice, and note it creates new transparency complications for professionals like lawyers.

Why it matters: This is one of the first concrete implementations of the EU AI Act's content-labeling rules by a major lab and could become a template other providers follow. The open question of whether watermarking measurably changes model output matters anywhere word choice is scrutinized, and connects to the broader run of AI-in-legal-filings controversies, including a recent case where an undisclosed AI-generated prompt injection was hidden in a court filing.