Anthropic, the company behind Claude, has implemented text watermarking to comply with the EU's Code of Practice on Transparency of AI-Generated Content. The watermarks are machine-readable but invisible to humans, embedded directly into the text by supported Claude models. When a model generates text, it weaves an imperceptible watermark that doesn't affect readability or meaning. These watermarks survive copying, pasting, and light editing, though they won't apply to very short texts or heavily modified content. All Claude models launched in the EU on or after August 2 automatically include this feature, regardless of whether users access Claude via API, Claude Code, Cowork, or Tag, or through providers like AWS, Google Cloud, and Microsoft Foundry. Anthropic is also working to extend watermarking to older models and plans a global rollout beyond the EU. The company is developing detection tools to identify these watermarks, which will be released soon. Notably, watermarks apply to all Claude output, even when the AI assists with translation, summarization, formatting, or proofreading, despite the Code of Practice exempting such assistive uses. Additionally, AI-generated images will use C2PA-based provenance metadata instead of easily removable visible watermarks. This move helps consumers identify AI content while also preventing AI models from training on other AI-generated text.

Source: GSMArena