A MULTIDIMENSIONAL CORPUS-BASED FRAMEWORK FOR ANALYZING LINGUISTIC COMPLEXITY AND STYLISTIC VARIATION IN HUMAN AND AI-GENERATED ACADEMIC WRITING

Abstract

The development of large language models (LLMs) has drastically altered academic writing. Systems currently exist that can write scholarly prose with context, grammar, and coherence. While these systems can write at a level similar to that of human scholars, important issues remain regarding the linguistic complexity and discourse of these systems, as well as the range and type of styles they can write in. Prior studies on issues of lexical diversity, syntactic complexity, or formulaic language tend to investigate these separate factors rather than considering academic writing as a multifaceted linguistic phenomenon. This study aims to develop a multidimensional system for reviewing academic writing to compare human and AI writing. The system will utilize lexical diversity, lexical sophistication, syntactic complexity, discourse cohesion, metadiscourse, readability, information density, and stylistic variation, among other features. Two corpora, one written by humans and one written by AI, will be constructed for review. The Human Academic Writing Corpus will be constructed with research articles, and the AI Academic Writing Corpus will be constructed using texts generated by AI in a restricted writing environment. The main theory that this study is grounded in is Biber’s (1988) work, although other theory and measures of lexical diversity, syntactic complexity, and metadiscourse will be integrated. This study aims to provide a framework beyond simple AI detection that considers the range of linguistic features in which human and machine academic writing converge and diverge. The framework is built to be used and researched in corpus linguistics, computational stylistics, English for Academic Purposes, academic discourse studies, and research related to generative AI.