Anthropic to Embed Invisible Watermarks in Claude AI Text

Anthropic has signed the EU AI Act Code of Practice and committed to marking new Claude models with machine-readable labels from August 2, 2026, globally across all Claude products, including API, Claude, Claude Code, Claude Cowork, and Claude Tag. Generated text will carry embedded watermarks, and generated files will have signed provenance metadata. The company also plans to provide detection tools for users and third parties. Two methods are used: invisible text watermarks that survive copying and minor edits, and signed metadata according to the C2PA standard for files like .svg, .png, .jpg. Cloud partners such as AWS, Google Cloud, and Microsoft Foundry will support text watermarks, but signed metadata may not be available everywhere. Anthropic notes limitations: a detected watermark does not guarantee Claude is the original author, as users often use Claude for editing, translation, or summarization; similarly, absent watermark does not mean content is not AI-generated due to heavy editing, paraphrasing, or format conversion. The effectiveness depends on robustness and detection tools. OpenAI reportedly has a 99.9% accurate text detector but hasn't released it, while Google DeepMind has open-sourced SynthID for text and integrated it into Gemini models. The decision may impact Claude's appeal to students, as detection could make AI use more transparent.