To comply with the European Union’s AI Act, Anthropic is implementing machine-readable watermarking across all new Claude models globally. Unlike the EU's narrow requirements, Anthropic is adopting a broad approach by marking all processed content, including minor edits like grammar corrections, even though the law provides exemptions for assistive editing that does not substantially alter meaning. For text, Anthropic will embed invisible watermarks by biasing word choices in detectable patterns, while non-text outputs will utilize C2PA metadata for provenance. Although the company plans to release a text detection API and technical details to meet EU transparency mandates, the system faces significant technical limitations. The watermarks can be neutralized by pasting text into other LLMs, through metadata editing, or via simple screenshots for media. Furthermore, the marks do not distinguish between entirely AI-generated text and human-authored text that has merely been processed by Claude. While the EU AI Act carries heavy penalties for non-compliance—up to 15 million euros or 3% of annual global revenue—the current implementation may result in ambiguous signals, as a detected mark only indicates that content "may have been processed by Claude" rather than confirming it is entirely synthetic.