Skip to article
← INTELLIGENCE DESK

Anthropic’s watermark plan shows regulation moving inside the model

New Claude models will mark generated text and attach provenance information to files as European transparency obligations take effect.

EDITOR'S READ

AI disclosure is becoming an infrastructure feature, not a label publishers can add only at the end.

Anthropic says models launched after the European Union’s transparency rules take effect will mark generated text and add provenance data to supported image files. The move follows Article 50 obligations designed to make synthetic or manipulated content identifiable.

The important change is architectural. A visible notice can disappear when content is copied, while metadata can be stripped during export. Model-level watermarking attempts to preserve a statistical signal inside the output itself, though durability, false positives and independent detection remain open questions.

For businesses, disclosure can no longer be treated as a final publishing checkbox. Content systems will need to preserve provenance through editing, translation and distribution, and explain when human work has materially transformed an AI draft. Regulation is beginning to shape the generation stack itself.

SOURCING NOTE

This briefing summarizes reported facts and adds independent context. It does not reproduce the source article's wording or structure.