SAN FRANCISCO — Anthropic began embedding invisible watermarks into text generated by its Claude AI models on Aug. 2, implementing the change in response to the European Union's AI law, which mandates transparency signals for AI-generated content.
The watermarking system applies an imperceptible signal directly into Claude's output. Anthropic said the mark does not affect text quality or readability and is designed to persist through copy-paste operations and some levels of editing.
The watermark operates at the model level, appearing in text generated across all Claude interfaces — the direct chatbot, its application programming interface and tools including Claude Code.
User backlash intensified Tuesday. Many users voiced concerns about the detectability of AI use in their work. Anthropic has not offered an opt-out option.
Technical experts warn the marks may not withstand common development workflows. Paraphrasing, specific code formatting or other application pipelines could strip the embedded signal, raising questions about the watermark's reliability as an identifier of AI-generated content.
For developers relying on Claude's API, that fragility creates compliance risk: if standard content processing defeats the watermark, its value as a dependable identifier disappears.
The EU's AI law is pushing companies including Anthropic to build such systems, adding compliance costs and forcing product design changes for AI developers operating in European markets.
AI music platform Suno also announced it would mark tracks created on its platform, a decision that followed legal challenges over content origin.
