How Claude's watermark probably works

All of us have probably read about Claude's watermark this week.

Claude models launched from Aug 2, 2026 onward will watermark ALL generated text and attach C2PA-signed provenance metadata to files. Older models don't watermark yet, but that's coming.

What's NOT confirmed: how it works. The technical spec is still "forthcoming"

So I dug into the most likely mechanism from the published research any production text watermark almost certainly builds on (Kirchenbauer et al., Google's SynthID-Text). The full step-by-step is in the infographic below:

Three things stand out to me:

1. The watermark lives in the word choices themselves. No hidden characters, no metadata. That is why it survives copy-paste and fades only under heavy paraphrasing.

2. Detection needs two things: enough text and the key. A short snippet may not provide enough statistical evidence to say much. And until Anthropic releases public detection tooling, only the key-holder can verify anything.

3. A detected watermark does NOT mean “Claude wrote this”. Anthropic is explicit about this: it means Claude processed the content.

   You write your own report and ask Claude to proofread it. Your human work may now carry an AI signature. That makes every “AI-generated vs. human-generated” policy a lot more complicated.

We're about to see provenance-by-default meet real organizations - attribution, audit, compliance, and trust in digital content.

That story is just getting started.

What are your thoughts on it?