2026-08-11 / Signal #2
Claude Now Embeds Invisible Watermarks in Every Generated Word
“Claude just signed every word it writes with invisible ink—and you can’t see the signature.”
The Story
Anthropic is embedding machine-readable invisible watermarks directly into text from supported Claude models (plus C2PA metadata for files). The marks are designed to persist through copy-paste and some editing for provenance, driven by EU AI Act transparency rules, though heavy paraphrasing can remove them and they do not prove ultimate human authorship. Reactions include excitement about verifiable AI content and immediate talk of watermark removers.
Why It Matters
Concrete product behavior/change tied to law (EU rules) that alters media, creator economy, institutions of trust, and content identity; weird story of hidden signatures living inside plain text; visual potential in “invisible ink” metaphors or courtroom provenance drama; builder utility for transparent workflows and cat-and-mouse detection tools; smart mischievous tone on the arms race.
Evidence
Anthropic’s rollout for newer Claude models (EU launches from Aug 2, 2026, applied globally), multiple articles and HN threads in the last 24 hours discussing provenance, detection industry impact, and removers.
Sources
Daily scan: 2026-08-11