Claude's Watermark Degrades on Short, Edited, and Code Outputs — Complicating EU AI Act Compliance

Anthropic confirmed August 14 that Claude's watermark uses Google DeepMind's SynthID-Text, a probabilistic method whose signal degrades on short texts, heavily edited passages, and code. A complete rewrite removes it. The EU AI Act, effective August 2, requires reliable identification of AI-generated content. The planned detection API returns yes-or-no answers whose confidence varies with text length and modification — a gradient, not a stamp.