Anthropic announced it will embed invisible watermarks in Claude's outputs to comply with the EU AI Act, which mandates labeling AI-generated content. The watermarks apply to all processed text, even human-edited content, and will persist through edits but may be stripped by other systems. The EU law requires watermarks for AI outputs post-August 2, with a grace period until December 2026.

Anthropic's approach, while technically compliant, risks over-labeling, as watermarks may appear on text lightly edited by Claude rather than fully AI-generated. Critics argue the system is easily bypassed and could confuse users about content origins. The company plans to release detection tools to meet EU requirements, though bad actors could remove watermarks via metadata editing.

The EU's transparency goals aim to combat misinformation, but the watermark's ambiguity may undermine trust. Anthropic's system may flag more content than legally required, including proofread or summarized text. The EU AI Act also imposes conflicting labeling rules, where AI-generated content may not need public labels if edited by humans.

The watermark's effectiveness and societal impact remain uncertain, with potential penalties for non-compliance.