Define the mistake first
A false positive occurs when human writing is classified as generated. The rate depends on the documents tested and the threshold used to make that classification.
A benchmark should explain both. Without that context, a precise percentage can be difficult to interpret.
Describe the human writing
Useful evaluations describe where the documents came from, how authorship was established, and which languages and genres are represented.
A sample of long essays answers a different question from a collection of short reviews. Results should make those differences visible.
Read the error rates together
Reducing one kind of error can change another. Examine false positives alongside missed generated writing, with the same decision threshold and a clearly described test set.
Editorial preview for the Proof5 design prototype.