Anthropic adds invisible watermarks to Claude texts globally

Anthropic is rolling out invisible watermarks on texts generated by its AI assistant Claude, a move tied to the EU's AI Act transparency code but applied worldwide. The system embeds a statistical pattern in the model's token choices, imperceptible to readers but detectable by specialized tools, and travels with text when copied or pasted Source: infobae. For image files like PNG, JPG, or SVG, Claude will attach digitally signed C2PA provenance metadata Source: abces. The watermark uses a version of Google DeepMind's SynthID-Text and does not add latency or token costs Source: lavozar.

Key details

• The watermark is applied at the model level, covering Claude, Claude API, Claude Code, Claude Cowork, and Claude Tag, as well as access via AWS, Google Cloud, and Microsoft Foundry Source: euronewses.
• It is not a definitive proof of authorship: a detected mark indicates Claude processed the content, but its absence does not prove human authorship Source: elcaribedr.
• The mark can be weakened or removed by heavy editing, paraphrasing, translation, or very short passages Source: lavanguardia.

Industry context

Anthropic joins OpenAI, Google, Meta, Microsoft, and Mistral in signing the EU's transparency code, while xAI and Chinese firms have not yet adhered Source: elespanol. The move responds to growing concerns about undisclosed AI-generated content, including in publishing, where editors report receiving AI-written manuscripts Source: elconfidencial. Anthropic plans to release a detection tool in coming weeks Source: lavozar.