Anthropic Details How Claude Marks AI-Generated Content: Invisible Text Watermarks and C2PA Provenance

Anthropic published a support document explaining how Claude marks the content it generates. The gist: AI-generated text and images will now carry traces that are invisible to the eye but detectable.

In brief

• What's changing: Claude models released on or after August 2, 2026 will mark their generated content at launch; work is in progress for older models. The basis is the EU AI Act's Article 50(2) Code of Practice on transparency, applied worldwide.
• Two methods: (1) an invisible watermark in text — embedded into the text itself, surviving copy-paste and some editing; it applies to all generated text. (2) signed provenance metadata — following the Coalition for Content Provenance and Authenticity (C2PA) open standard; for images it covers .svg, .png, and .jpg files.
• Where it applies: Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag; also AWS, Google Cloud, and Microsoft Foundry (with some limitations).
• Limitations (as Anthropic itself notes): a detected mark indicates content "may have been processed by Claude" — it's not full, conclusive proof. Marks may not survive heavy editing, paraphrasing, or format conversion; very short text may be too minimal to detect reliably. Anthropic says it will let users and third parties detect these marks, with technical documentation to come.

Our take

This is "transparency infrastructure" trying to keep pace with regulation (the EU AI Act) — and it's directly relevant to our audience, since most of us use AI when producing content. The practical lesson runs two ways: first, the assumption that "AI-made content is undetectable" no longer holds; text and images increasingly carry invisible marks and provenance. Second, the mark isn't perfect — heavy editing can strip it, it isn't conclusive proof, and the detection tools aren't public yet. The bigger picture: documenting where content comes from is slowly becoming standard. Wherever AI and human work mix, your own labeling and transparency habits will matter more — both for trust and, where rules like the EU's apply, for compliance. This isn't something to dodge; it's a norm settling into place.

Kaynak: Anthropic Support