AI 87% is not a measurement: TOEFL essays hit 61% false positives

The piece opens with a hypothetical: a teacher sees "AI 87%" on a student essay and reads it as a measurement. The author's point is that it is a proxy classifier's output, not a reading of authorship. Every detector maps a document to a score and flags it above a threshold, but it never observes the true label — only surface text — so what it tests is whatever property happened to differ between human and machine writing in its tuning data.

The evidence the author leans on is an out-of-distribution test. Liang et al. (2023, Patterns) ran seven widely used detectors against 91 human-written TOEFL essays and 88 US eighth-grade essays, and the TOEFL essays averaged around 61% false positives. The author then highlights the intervention: simplifying native-written vocabulary toward the second-language distribution pushed misclassification sharply higher, showing the label follows the confounder. The author also notes the study's own limits — a small pilot sample, and detectors from early 2023 that largely leaned on older, smaller language models.

From there the piece widens to operations: one displayed percentage hides where errors cluster, and detector disagreement itself carries information about uncertainty. The design checklist asks for local evidence, group-specific error rates, uncertainty ranges, and a human in the loop — framing a detector result as one piece of evidence rather than proof, and the final judgment as the human's.