Claude embeds invisible watermark in all new models amid backlash

Anthropic has signed the EU AI Act's Code of Conduct for AI-Generated Content Transparency, which mandates that new models released after August 2, 2026, must embed invisible watermarks in generated text. The watermark is not metadata but becomes part of the text, making it copy-paste resistant and surviving some editing. This mechanism will be applied globally, not just in the EU, and will cover all Claude products including API, Claude, Claude Code, Claude Cowork, and Claude Tag, as well as cloud customers via AWS, Google Cloud, and Microsoft Foundry. Additionally, Anthropic will add signed metadata to supported file types like SVG, PNG, and JPG following the C2PA open standard. The company is developing detection tools for users and third parties to verify if content was processed by Claude but acknowledges limitations: detecting a mark does not confirm authorship, and absence of a mark does not guarantee non-AI origin. The algorithm details are not public. Earlier, a Reddit user discovered that Claude Code contained hidden Unicode characters that marked Chinese users, suggesting Anthropic has used such techniques before. The article expresses skepticism about the watermarks not affecting generation quality.