Claude Begins Secretly Watermarking AI Content — Marks Survive File Downloads

Anthropic has announced it will mark content generated by Claude with invisible watermarks and digital metadata. The embedded text watermark remains even when copied from the AI chat into another document. All Claude models released after August 2, 2026 support the new marking. The watermark is machine-readable and invisible to users, detectable via special tools. It applies across all Claude services, including API, Claude Code, Claude Cowork, and Claude Tag. For media files in SVG, PNG, and JPG formats, Claude embeds metadata following the C2PA standard, indicating AI creation and allowing origin checks. Watermarking is global, including via AWS, Google Cloud, and Microsoft Foundry, though some third-party platforms may not support C2PA. Anthropic warns the system is not fully accurate: edited, translated, or shortened text may gain a watermark, and absence of a mark does not prove human authorship. Metadata can be stripped by format conversion, resaving, or screenshots. The move aligns with the EU AI Act's Code of Practice for AI content transparency, with older models getting the feature during a transition period.