What a detector would actually change
The most consequential fact about Claude's watermark is that nobody can read it yet, and the second most consequential is that this will not stay true.
The retroactive problem
Marking started on 2 August 2026. Detection has not started at all.
That gap does not protect anything. Text published today carries whatever mark it carries, and remains markable whenever the tooling appears. You are publishing into a future that can look backwards.
Why third parties cannot build one
A statistical watermark is detectable only with the key that seeded it. Without it there is no test to run — you are looking at ordinary words.
This is deliberate. A mark anyone could detect is a mark anyone could strip, and the design would defeat itself.
What it will and will not settle
A detector will report a likelihood, not a verdict. Anthropic already says a mark is not fully conclusive and does not establish authorship.
The risk is not that detection will be wrong. It is that a score will be treated as proof by people who did not read that far.