Notes on AI-text detection, false positives, error rates, and what a detector can and cannot tell you.
28 September 2026 · 7 min read
A detector can flag a human, and it can miss a machine. Both errors come from the same threshold, which means no setting makes both disappear. Why "near-zero error" claims describe one side of a dial, and how to choose which error you optimise against.
28 September 2026 · 7 min read
A Stanford study found 61.22% of real human TOEFL essays were labelled AI-generated, a bias multiplier of roughly 19×. Why it happens, why it is fixable, and the engineering choice we made about whose writing counts as human.
28 September 2026 · 6 min read
Every editor now fears two things: publishing AI-written work, and rejecting a human writer because a number said so. A four-step workflow for screening at a high threshold, and the four things you must not do with a statistical score.
28 September 2026 · 7 min read
Books written before ChatGPT existed come back as "29.7% AI". A professor's 1981 article scores 77%. The Declaration of Independence scores 99.99%. Here is why that happens, who gets hit hardest, and what a false-positive rate actually tells you.