Every Sentence Claude Writes Now Carries an Invisible Mark - and It Can't Tell Anyone Who Wrote It
About This Episode
Anthropic published a technical explainer on August 14 describing how it is embedding invisible, machine-readable watermarks in all text Claude produces, using a version of Google DeepMind's open-source SynthID-Text. The mark is being applied worldwide - not just in Europe - to satisfy EU AI Act transparency rules that took effect August 2. Anthropic is explicit that detecting the watermark only shows Claude was likely involved at some point, and cannot establish who actually wrote the text.
Our Take
Anthropic just shipped the first mass-scale verifier for machine-written text, and it checks the one thing that does not settle any argument - it proves a machine was involved, not who did the work, and it is weakest exactly where someone stands to be accused.
Continue Reading on Unscarcity
The Verifiability Premium: When Proof Replaces Credentials
The article argues that the decisive question for any profession is whether its output has a checker - and the watermark is the first mass-scale checker for AI text, except it verifies provenance rather than whether the work is any good.
GPT-4.5 Beats the Turing Test. Now What?
For seventy-five years the test for machine writing was whether a reader could tell; the watermark answers that by tagging the machine instead of testing the reader, and like the Turing test it delivers information other than the information people think they are getting.