Anthropic to Add Invisible Watermarks to Claude Outputs Globally for EU AI Act Compliance

Ars Technica AI · rss · 2026-08-13

To comply with the EU AI Act, Anthropic announced it will add machine-readable watermarks to content generated or processed by its models.

Global Rollout: All new models offered globally will feature the watermarks from day one, not just in the EU.
Scope: Text outputs will carry invisible embedded watermarks, while other generated files will include digitally signed provenance metadata where supported.
"Nuke it from Orbit" Approach: The EU guidance exempts minor assistive edits like grammar correction. However, model-level watermarking cannot distinguish between wholesale generation and a minor comma fix. Consequently, Claude will stamp all processed content, potentially capturing exactly what the law intended to leave alone.

Anthropic stated it will eventually release a detection tool and share technical details to meet legal requirements.

More from Models