Multilingual AI-Text Detection Performance and False-Positive Risk: A Statistical Analysis of WordBinary's Internal Benchmark

This paper presents a statistical analysis of WordBinary’s internal multilingual AI-text detection benchmark across 27 languages. The study examines reported accuracy, precision, recall, F1 score and human false-positive behaviour for Arabic, Chinese, Czech, Danish, Dutch, English, Finnish, French, German, Greek, Hindi, Hungarian, Indonesian, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Turkish, Ukrainian and Vietnamese. Across the reported languages, mean accuracy was 99.78%, mean precision was 99.65%, mean recall was 99.87% and mean F1 was 99.76%. English achieved 100.00% recall with a reported human false-positive rate of 0.59%, while the remaining 26 languages were reported as having a human false-positive rate near zero. The paper also examines the practical meaning of false-positive rates through base-rate and large-scale screening calculations. These analyses show that even a low false-positive rate can affect a meaningful number of genuine writers when large volumes of documents are screened, and that the predictive value of a positive result depends on the prevalence of detectable AI-generated writing in the evaluated population. The study argues that AI-detection results should be used as review signals rather than automatic proof of authorship or academic misconduct. It recommends combining detector output with highlighted-passage review, drafts, version history, research notes, source records, author explanations and applicable institutional policy. The reported performance figures are first-party WordBinary internal benchmark results and should not be interpreted as independent third-party validation. The paper identifies the additional methodological information required for reproducibility, external validation and subgroup fairness assessment.