Anthropic adopts broad watermarking for Claude to meet EU AI Act obligations
The move applies invisible, machine-readable tags to text and files, drawing criticism for potentially confusing users and failing to distinguish between wholesale generation and minor edits.
Anthropic has announced it will embed invisible, machine-readable watermarks in all content processed by its Claude models, a move designed to comply with the European Union’s AI Act. The implementation applies globally from day one for new models and extends to previously released models by December 2026. While the EU law exempts assistive editing functions, Anthropic is adopting a broad approach that marks any content touched by the AI, even if the underlying ideas are human-authored.
The company confirmed that text outputs will carry embedded watermarks invisible to the user, while other generated files will include digitally signed provenance metadata where supported. This “nuke it from orbit” strategy applies watermarks to all processed content, despite the EU AI Act not requiring labels for assistive functions such as grammar correction or standard editing that do not substantially alter meaning.
Critics note that the watermarks may be easily bypassed and could confuse users by failing to distinguish between wholesale generation and minor edits. Text watermarks work by biasing the model’s word choices in a pattern spread across the entire document, which may result in slightly suboptimal word choices to maintain the signal. Furthermore, the watermarks may be destroyed if text is pasted into another chatbot system that edits the content.
Anthropic acknowledged that the watermark may appear on content that was not generated by Claude, such as human-authored text simply edited in a workflow involving Claude. The company stated that a detected mark provides a signal that content was processed by Claude but is not fully conclusive. This approach contrasts with EU guidelines that aim to ensure AI tools do not upset the integrity and trust in the information ecosystem.
The EU AI Act imposes fines of up to 15 million euros or 3% of worldwide annual revenue for violations. Ars Technica contacted Anthropic for a timeline on detection tool release and testing results, but received no immediate response. The company plans to release detection tools to support transparency, though the effectiveness and reliability of these tools remain to be seen.
