Detection Won't Save You

Anthropic says future Claude models will watermark text. Not visibly. Not with metadata. A statistical pattern in the word choices that can later show Claude was probably involved.

The easy read is compliance. The useful read is workflow.

A watermark can tell you a model touched a piece of text. It cannot tell you who asked for it, what source material it used, who edited it, whether legal saw it, or why it ended up in the CMS without a disclosure note.

That is the part companies keep trying to skip. They want detection to behave like governance. It does not. Detection happens after the work escapes. Governance has to live before that: in intake, review, asset records, approval queues, and handoffs between tools.

Every customer-facing content path now needs provenance attached to the asset. Which system touched it. What it did. Who reviewed it. What label, if any, belongs on the final thing.

The watermark is not the control. It is a flare. If it tells you something you did not already know about your own content, the failure happened upstream.