Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

the watermark followed the sentence

Anthropic says supported new Claude models will embed watermarks in generated text worldwide, before public detection tools are available.

Published Updated Read 2 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-08-12

a Claude sentence may soon carry more than its visible words. Anthropic says text from supported new models will include an imperceptible watermark that survives copying and may survive editing; the promised detector is still on the way.

1. the sentence acquired a passenger

sources:

what happened:

Anthropic has signed the EU AI Act’s Article 50(2) transparency code and says Claude models launched in the EU on or after 2 August 2026 will support machine-readable marking at launch. For supported models, the company says generated text gets an embedded watermark, while supported files can receive signed provenance metadata. It says the marks will apply across its API, Claude, Claude Code, Cowork, Tag, and cloud partners wherever the supported models are offered; older models are still being brought into the scheme. Anthropic also says a detector for users and third parties is forthcoming, not available yet.

the collision:

This is not a badge on a chat screen. It turns ordinary copied text into a potential provenance surface controlled at the model layer. That can give a reader useful context, but it also changes the custody problem: a sentence can travel into a document, a note, or someone else’s draft with a hidden claim attached, while the person handling it cannot yet independently check what the claim says or how it survived the trip.

question left open:

Before invisible provenance becomes normal plumbing, what should a person be able to inspect, remove, preserve, or contest in the text they are responsible for publishing?

left on the table

  • Open Science Desktop is a serious local-first research workbench with reproducible run records, but its latest movement is product polish rather than a distinct public shift today.
  • Anthropic’s public skills repository remains a large distribution surface, not new evidence that a person gained a clearer boundary over what an installed skill can do.
  • OpenCode memory was cut: another local-memory object one day after the local-agent memory-budget fieldnote is the same week wearing a different hat.
provenance transparency personal-ai