DETECTION METHOD LANGUAGE AND PLATFORM CHARACTERISTICS PREDICT MACHINE GENERATED TEXT IDENTIFICATION ACROSS MULTILINGUAL SOCIAL PLATFORMS

This analysis examines whether detector rankings remain consistent across multilingual social media environments when evaluation coverage is standardized. A descriptive secondary analysis uses 1,032 published Multi Social evaluations from 2025, covering twelve detectors, twenty-two languages and five platforms. Comparisons summarize AUROC distributions and match detector, language and platform groups. Configurations adapted to Multi Social achieve mean AUROC of 0.9622, compared with 0.5470 for previously trained alternatives, with higher family averages in all 86 common evaluation combinations. Across nine languages with complete platform coverage, the spread in language means narrows from 0.2116 among previously trained configurations to 0.0171 among adapted configurations. Platform ordering also changes: Telegram leads unrestricted comparisons, whereas Discord leads after matching language coverage. This reversal demonstrates how evaluation composition shapes aggregate rankings without changing underlying scores. The analysis refines the descriptive application of domain adaptation reasoning by distinguishing consistent configuration advantages from composition sensitive comparisons. Detector procurement and governance should require matched evaluations and documented linguistic coverage before translating pooled benchmark rankings into operational decisions.