AI text detectors are increasingly deployed in academic contexts, yet it remains unclear whether their outputs truly reflect AI‑generated authorship or merely capture the polished academic English style. Prior work reported high false‑positive rates (FPR) for non‑native English writing, but those studies often confounded authorship with differences in topic, discipline, and writing style.
Professional editing services offer a natural experiment: the same manuscript is rewritten for linguistic fluency while preserving the original author and content. We gathered 135,389 manuscript pairs from a professional English editing service spanning 2018‑2025, and evaluated 13 widely used AI text detectors on both the original non‑native versions and their native‑edited counterparts.
The detectors exhibited a striking range of human‑text FPRs, from 0.0% up to 100.0%. Moreover, the same edits could raise AI scores in some detectors while lowering them in others. Score changes correlated positively with the magnitude of editing, indicating that more extensive revisions produced larger score swings.
These findings identify professional editing style as a primary confounding factor in AI detector outputs, rather than a clean separation between text origin and linguistic style. The results raise concerns about fairness and reliability when such tools are applied in academic evaluation.
Blogger's Review: By leveraging a massive, content‑controlled before‑and‑after editing dataset, the study compellingly demonstrates how writing style alone can drive AI detector outcomes. It serves as a cautionary reminder that reliance on these tools—especially for non‑native authors—must account for editorial influences to avoid systematic bias.