Studies in Linguistics and Language Education

Constructing Academic Voice: A Corpus-based Comparison Of Human- And Ai-generated Doctoral Thesis Conclusions

Auteurs : Ayachi Dhaya Eddine .

Résumé

The current study examines the linguistic and interpersonal features of AI-generated academic writing by comparing human-written and AI-generated general conclusions in Applied Linguistics PhD theses. While most previous research has focused on ethical concerns and pedagogical impacts, limited attention has been given to how AI constructs stance and academic voice. Using a corpus-based approach, the study analyzed two corpora: a Human Corpus (HC) containing 40 authentic general conclusions extracted from open-access theses and an AI Corpus (AIC) consisting of 40 conclusions generated using Google’s Gemini 3 Flash Preview. The AI-generated conclusions were produced by providing the model with theses without their original conclusion sections. The analysis adopted Hyland’s (2005) stance model, focusing on hedges, boosters, attitude markers, and self-mentions. Quantitative methods examined marker frequencies, while qualitative analysis looked at their contextual use. Results indicated that human-written conclusions were nearly twice as long as AI-generated ones, likely due to model input/output limits and prompt instructions that may influence conclusion length. AI texts tend to overuse boosters and self-mentions, resulting in a more assertive and less balanced academic tone, whereas human writers typically use more hedges, reflecting greater epistemic caution. No major difference was found in the use of attitude markers. Qualitative analysis further showed that AI writing tended to be formulaic, relying on expressions such as “This study” and “The present study”, while human writers demonstrated greater lexical diversity, context-specific wording, and a clearer authorial voice through the use of personal pronouns such as “I” and “we”.

Mots clés

AI-generated academic writing; Stance markers; Academic voice; Corpus-based analysis; Applied Linguistics; Hyland’s stance model; Language Models (LLMs)