Lexico-syntactic complexity in human and AI-generated outputs across interpreting and translation: a paired-corpus study

This study examines differences in lexico-syntactic complexity between human translation outputs and AI-generated texts under interpreting and translation conditions, with implications for computer-assisted language learning (CALL) and translator/interpreter education. The dataset comprises 60 source segments and their corresponding human and AI outputs, yielding 120 texts. Human and AI texts were source-paired within each mode, while cross-mode comparisons were conducted using quasi-comparable corpora. Five lexical and eight syntactic complexity measures were analyzed to examine mode-related and agent-related differences. The results show dimension-specific mode dependence: AI-human differences in lexical complexity varied significantly across modes, whereas syntactic differences did not. Under the interpreting condition, AI-generated texts showed higher values on all five lexical measures, whereas human interpreting showed higher values on five syntactic measures. Under the translation condition, AI-human differences were reduced and confined to lexical density and sophistication, with no statistically significant differences on the eight syntactic measures. These findings suggest that AI-generated texts tend to display greater lexical concentration and written-like features, while human interpreting reflects characteristics of online language production and structural extension. Source-paired human and AI texts can serve as comparative resources for CALL and translator/interpreter training, helping learners critically evaluate AI outputs through linguistic features, textual evidence, and production conditions. Overall, AI-human complexity differences are not fixed properties of translation agents but vary according to complexity dimensions and production contexts.