The Pen is Mightier Than the Algorithm? A Multilevel Linguistic Comparison of LLM‐ and Human‐Translated Research Article Abstracts

ABSTRACT Translating local research into English as a lingua franca (ELF) connects local scholarship with global readership, but this process remains constrained by language barriers. Large language models (LLMs) offer advanced accessible solutions, but their responsible integration into academic translation requires a deeper understanding of the linguistic profiles they produce. To address this gap, this corpus‐based study quantitatively evaluates three linguistic dimensions of research article abstracts (RAAs), i.e., lexical complexity, syntactic complexity, and cohesive features, that can influence knowledge dissemination of research articles. Using Chinese‐to‐English RAAs from soft and hard disciplines, we compare translations by two LLMs (LLMTs), GPT‐4o and DeepSeek‐V3, against their corresponding de facto languaging practice by human translators (HTs). Findings reveal that both LLMTs consistently produce higher lexical complexity through varied and sophisticated vocabulary beyond academic stylistic norms but feature lower syntactic complexity with less subordination, reduced phrasal complexity, and weaker cohesive strength. HTs, however, balance lower lexical complexity with higher syntactic complexity and stronger cohesive ties. These findings highlight the communicative affordances and constraints of LLM‐mediated academic translation, offering practical and pedagogical insights for English for Academic Purposes practitioners, academic translators, and non‐anglophone scholars in ELF academic contexts.