Anthropic now watermarks all Claude text and files, worldwide
An imperceptible signal hides in the text, C2PA tags the files, and it spans the API, Claude Code and Cowork — with no opt-out and no way to prove much.
Copy markdown
Two marks: a hidden signal in text, C2PA in files
Claude weaves an imperceptible watermark straight into generated text — it survives copy-paste — and signs .png, .jpg and .svg files with C2PA provenance metadata. Two separate systems, both on by default.
It's on everywhere you touch Claude — the API included
Marking spans the Claude apps, Claude Code, Claude Cowork, Claude Tag, the API, and cloud deploys on AWS, Google Cloud and Microsoft Foundry. Models launched Aug 2 or later mark at launch; older ones are being retrofitted.
No opt-out, and a hit isn't proof
Anthropic documents no switch to turn it off. And detection is probabilistic: heavy editing, paraphrasing, translation, screenshots or format conversion can strip the mark, and very short passages may carry nothing.
Why now: the EU AI Act, shipped globally
The trigger is the EU AI Act's Aug 2 transparency rules for AI content, but Anthropic rolled marking out worldwide rather than region-gating it. If you ship Claude-generated text or images, assume they now carry a fingerprint.
Elsewhere: 60 Claude subagents nudged a 160-year-old math bound
Anthropic says an unreleased Claude, running ~60 subagents across two sessions and 31M output tokens, pushed a Riemann zeta lower bound from 41.6% to 67.2% — a concrete high-water mark for what big parallel agent runs can now attempt.