Topic: what the groups write about, and effects with topic held fixed

How to read the cards, the status badges and the four groups is explained on the Comparisons overview; every statistical term links to the Statistical glossary.

What the groups write about

The populations differ in what they write about as well as how. Each publication carries a category from the classifier described on the Methods page (one of Substack's 33 leaderboard categories; 64% exact and 81% top-two agreement with Substack's own placement on a blind check), so a stratum below is "publications the classifier put in Business", not a ground-truth topic. The topic selector on the Posts and Text features pages redraws every post-level card for one category.

Share of posts labelled AI, by category

Share of scanned posts labelled Fully AI-assisted text, with 95% , for categories with at least posts. Posts were sampled uniformly, so these are shares of posts, not of publications.
The same, for any label other than Fully Human-Written.

Holding topic fixed

Every post-level variable's human-versus-AI δ recomputed with topic held fixed (): pairs of posts whose publications share a category, each category weighted by its number of pairs, with 95% bootstrap intervals. Compare with the overall δ in the "At a glance" chart on the Comparisons overview: where this value is much smaller, the overall difference was largely topic mix. Exploratory, outside the correction family.