Download the free AI Labelling + Watermarking Field Guide:
Get AI-Ready with Kyle’s 5-Day Challenge:
Subscribe and turn on notifications to catch the next live stream:
Anthropic says new Claude models launched in the EU must support machine-readable marking. That means embedded watermarks for generated text and signed provenance metadata for supported files…but the viral claim that every Claude response is already publicly detectable is too broad. There’s a lot of bad info going around.
I break down what Anthropic has actually committed to, how text watermarking differs from C2PA metadata, why a detected mark does not prove Claude authored the work, and why no mark does not prove a human wrote it. The technical documentation and public detection tools are still coming, so anyone claiming certainty about the exact implementation is getting ahead of the evidence.
—— Time Stamps ——
0:00 Claude’s Hidden Watermarks
1:15 What Anthropic Actually Said
2:15 What the EU Rule Requires
3:30 Text Watermarks vs File Metadata
4:56 How C2PA Provenance Works
6:34 Why Copy-Paste May Not Remove It
8:10 How Text Watermark Detection Works
10:18 Why Detection Can Be Wrong
12:41 No Mark Does Not Mean Human
14:10 The Detector Is Not Ready
15:07 Can Watermarks Be Removed?
16:57 Claude Is Not the Only AI Affected
19:07 What This Actually Means
—— Useful Resources ——
Anthropic’s transparency commitments:
EU AI Act information:
C2PA specification:
Find everything else at
コメント