The new marking system is a provenance signal, not a plagiarism detector — and the HN discussion shows how fast that distinction collapses in practice.
- 🔍 Text gets an imperceptible watermark that survives copy-paste; files carry C2PA signed provenance metadata for .svg/.png/.jpg.
- 🎯 It answers “did this pass through Claude,” not “did Claude write this” — proofreading or translating your own draft leaves the same mark.
- ⚠️ Weak positive, no real negative: heavy editing, translation, or blending erases the mark, so absence proves nothing.
- 📊 Coverage is model-gated — models launched on or after Aug 2, 2026, across the API, Claude Code, and the apps — which matters if you build on the platform.
- ⚡ The detection mechanism isn’t public yet (“forthcoming technical documentation”), so you’re trusting a signal you can’t yet audit.
- 💡 For enterprise governance it’s a compliance primitive, not an adjudication tool — treat a hit as “investigate,” never as a verdict.
The reaction across the web tracks that gap almost exactly. Notebookcheck is openly skeptical the scheme works, since a mark only shows text passed through Claude rather than that Claude authored it, and The Cryptonomist notes paraphrasing quietly erases it. The measured takes land on the same worry: explainX accepts the EU-AI-Act compliance motive but warns institutions will read a probabilistic signal as a verdict, repeating the AI-detector era’s mistakes, while The Decoder credits the open C2PA choice yet stresses the same authorship caveat. The trend isn’t “watermarking bad” — it’s that a checkable-at-scale provenance mark keeps getting sold as the authorship proof it was never designed to be.