Anthropic's Claude leaves invisible AI watermarks globally
Claude 3.x now embeds hidden AI watermarks in text and files worldwide...
Anthropic has implemented a global marking system for AI-generated content produced by its Claude models, effective August 2, 2026, to comply with new EU AI Act transparency requirements. The system embeds two types of markers: imperceptible watermarks in generated text that remain detectable by compatible systems, and digitally signed C2PA provenance metadata in file outputs that creates a tamper-evident record.
However, the durability of these markers is inconsistent outside the Claude ecosystem. While copying marked text may preserve the watermark, substantial rewriting, translation, or editing of files—especially when metadata is stripped—can render detection impossible. Anthropic emphasizes that a detected marker confirms Claude processed the content but doesn’t guarantee authorship of every word, while the absence of a marker doesn’t prove human origin.
- Claude 3.x models now embed invisible watermarks in text and C2PA provenance metadata in files globally, complying with EU AI Act transparency rules.
- Watermark durability drops with editing—translation or rewriting can break detection, while file provenance depends on preserved metadata.
- Detected marks indicate AI processing but not full authorship; organizations need policies for AI disclosure and disputed claims.
Why It Matters
AI provenance tracking forces businesses to reassess content authenticity policies and disclosure processes in regulated workflows.