anthropic's invisible watermarks can survive copy-paste and editing
anthropic will embed c2pa metadata and imperceptible text watermarks in all claude outputs to comply with the eu ai act. this machine-readable tagging persists across copying and editing, setting a new standard for ai provenance.
anthropic is implementing a multi-layered provenance system for claude, embedding c2pa metadata in images and imperceptible watermarks directly into text generation. these markers are applied at the model level, meaning they remain intact across the api, aws, google cloud, and third-party platforms, and survive standard copy-paste actions. the compliance move targets the eu ai act's four-month grace period for existing models, prioritizing detection for future releases while retrofitting current ones. by enabling third-party detection, anthropic aims to help platforms and users identify ai-generated content, addressing growing concerns about misinformation and synthetic media.