Too Much of One Thing, Too Little of Another? Comparing Cohesive Networks Between Humans and ChatGPT in Academic Writings and Translations

Abstract

ABSTRACT Artificial intelligence (AI) is increasingly recognised for its potential to support academic writing processes, sparking growing interest in the comparison between human and AI writings. Nevertheless, much existing research assumes AI functions as an autonomous academic author, which may not fully reflect real‐world academic practices. This study provides a more realistic perspective into the human‐AI distinction in language output by comparing textual cohesion in academic texts translated by human translators and Chat Generative Pre‐trained Transformer (ChatGPT), using original English academic writing as a benchmark. Combining Least Absolute Shrinkage and Selection Operator (LASSO) regression and random forest modelling, this article reveals that human‐translated academic texts are characterised by the overuse of additive conjunctions, whereas the most distinctive feature of ChatGPT translations is the underuse of determiners. Compared to original academic writing, both translation varieties display explicitation tendencies, manifested in the increased use of additive connectors and third‐person plural pronouns. Yet, despite these features of explicitation, both human and ChatGPT translations exhibit lower levels of logical coherence and semantic similarity between sentences compared to original academic texts. These patterns highlight the need to enhance learners’ awareness of logical structuring and semantic clarity, particularly when translating from their first‐language writings into academic English.

Ähnliche Arbeiten