Metadiscursive nouns in academic argument: ChatGPT vs student practices

The ability of ChatGPT to create grammatically accurate and coherent texts has generated considerable anxiety among those concerned that students might use such large language models (LLMs) to write their assignments. The extent to which LLMs can mimic human writers is starting to be explored, but we know little about their ability to use nominal resources to create effective academic texts. This study investigates metadiscursive nouns in argumentative essays, comparing how ChatGPT and university students employ these devices to organise text, express stance, and construct persuasive arguments. By analysing 145 essays from each source, we examine the syntactic patterns, interactive functions, and interactional uses of metadiscursive nouns. The analysis reveals that while overall frequencies were similar, ChatGPT has distinct preferences for simpler syntactic constructions (particularly the determiner + N pattern) and relies heavily on anaphoric references, whereas students demonstrate more balanced syntactic distribution and greater use of cataphoric references. Interactionally, ChatGPT prefers manner nouns for descriptive precision, while students favour status nouns for evaluative reasoning and evidential nouns for empirical grounding. These findings show that, while structurally coherent, LLM-generated texts often lack the rhetorical flexibility and evaluative sophistication of human academic writing, offering valuable insights for EAP pedagogy.

Introduction

The emergence of large language models (LLMs) such as ChatGPT has triggered significant alarm about their potential to impact academic writing practices, presenting both opportunities and challenges for English for Academic Purposes (EAP). These AI-powered models demonstrate remarkable capabilities in generating human-like texts across diverse academic genres, employing sophisticated natural language processing algorithms to produce coherent written discourse (OpenAI, 2024). However, it is crucial to acknowledge that these models do not engage in human-like cognitive processes nor possess genuine understanding, intention, or awareness (Byrd, 2023; Gallagher, 2023). The apparent “intelligence” demonstrated by LLMs is, in fact, an emergent property derived from statistical patterns identified within their training datasets, rather than a manifestation of human-like reasoning or comprehension. Consequently, in this study, we adopt the term “LLM” to refer to these computational models, including systems like ChatGPT.

While initial enthusiasm focused on ChatGPT's potential to support language learning and writing development, deeper questions have emerged about its impact on academic integrity and writing pedagogy. Proponents highlight its potential to scaffold students' writing development, provide instant feedback, and explain complex language concepts (Johannesson, 2024; Su et al., 2023). However, critics raise concerns about the challenges of distinguishing LLM-generated from human-written texts, with implications for academic integrity and assessment (Herbold et al., 2023; Ingley & Pack, 2023). Despite the development of detection tools like GPTZero and AICheatCheck, reliable identification of LLM-generated texts remains problematic (Nguyen & Barrot, 2024; Scarfe et al., 2024), prompting researchers to investigate the distinctive linguistic features that characterise LLM-generated discourse. Recent studies of ChatGPT-generated texts have primarily focused on lexical and grammatical features, suggesting these texts tend to be “vaguer and more formulaic” (Gao et al., 2023, p. 1) and sometimes “empty or fluffy” (Markey et al., 2024, p. 22). However, the need to examine how LLM-generated texts deploy the sophisticated rhetorical resources that characterise effective academic argument remains urgent. We address this gap by investigating the use of metadiscursive nouns, a type of unspecific abstract noun like argument, analysis, and finding. These serve as rhetorical devices in academic writing to both organise discourse and express authorial perspective on propositional material (Jiang, 2022; Jiang & Hyland, 2018).

Examining both the syntactic patterns and rhetorical functions of these nouns, we set out to determine what differentiates LLM-generated from human-written texts in their use. To explore this feature, we prompted ChatGPT to produce argumentative essays on identical topics and comparable lengths as those written by British university students, enabling a systematic comparison between them. Our analysis contributes to understanding automated text production and its implications for writing pedagogy and assessment in EAP.

Metadiscursive nouns: interaction and persuasion in academic writing The term “metadiscursive noun” was first used by Francis (1986) and according to Jiang and Hyland (2018), they are a type of unspecific abstract noun that “refer to the organisation of the discourse or the readers’ understanding of it” (p. 510). Their role in constructing the textual interaction is achieved through “contextual lexicalisation” (Jiang, 2022, p. 33), wherein their specific pragmatic meaning is derived from material within the immediate context. This assists writers to reference

Metadiscursive nouns used by ChatGPT and British students: an overview We identified 349 metadiscursive nouns in ChatGPT outputs (4.79 per 1000 words) compared to 422 instances in student essays (5.41 per 1000 words). The difference is not significant, which surprised as earlier research shows notably lower frequency of metadiscourse in ChatGPT essays (Jiang & Hyland, 2024b). The similar frequencies, however, can be explained by several studies which have shown ChatGPT's propensity for noun-based features either in academic prose, including noun-related bundles (

Interactive use of metadiscursive nouns by ChatGPT and students As discussed in Section 3, a key aspect textual interaction that writers construct using metadiscursive nouns is the creation of cohesive and coherent discourse organisation. This is achieved by either pointing backward to previously discussed material or forward to upcoming propositions, aligning with what writers assume readers expect to find and where they may need textual support. Such anaphoric and cataphoric references are closely linked to specific syntactic patterns of metadiscursive