Anthropic announced plans to embed invisible machine-readable watermarks in Claude-generated text, complying with the European Union's AI Act transparency requirements. The watermarks are imperceptible to humans but designed to help both people and platforms identify AI-generated content. The company will deploy them globally across Claude Platform via API, Claude, Claude Code, Claude Cowork, and Claude Tag. New models will ship with watermarking enabled by default, while support for older models is still in development.
The watermarking approach uses two distinct methods. For images, Anthropic adopted the C2PA standard already used by Adobe, OpenAI, and Google. For text, the watermark is applied at the model level itself, designed to persist even when content is copied, pasted elsewhere, or lightly edited. Anthropic emphasizes the process doesn't alter meaning, technical quality, or readability.
Within hours of launch, developer Guillaume Meyer released Watermarks Remover on GitHub—an open-source tool specifically designed to strip these protections. The project operates in two layers: Layer A identifies and removes invisible Unicode characters and markers embedded directly in text, while Layer B weakens statistical watermarks through aggressive rewriting techniques like paraphrasing and back-translation. The tool also removes C2PA, XMP, and other metadata from PNG, JPEG, SVG, PDF, DOCX, ODT, HTML, and Markdown files.
Meyer acknowledges that removing statistical watermarks is "the best possible" approach, noting that without Anthropic's published detector and keys, the tool cannot guarantee officially undetectable output. Still, the rapid emergence of a bypass tool highlights the technical challenge of creating watermarks that truly resist modification.