Decoding the Machine: A Corpus-Based Analysis of Pragmatic Structures in AI-Generated Discourse and Their Interpretive Implications for EFL Learners

Abstract

Conversational AI systems such as ChatGPT increasingly function as a primary source of linguistic input for EFL learners, yet existing research has examined either AI's capacity to generate pragmatically appropriate language or AI-assisted learning outcomes, leaving largely unexplored how the structural features of AI-generated responses may themselves create interpretive difficulty. This study addresses that gap through a corpus-based analysis of AI-generated responses from the WildChat dataset. A corpus of 122 human-AI exchanges (36,233 word tokens), sampled across four pragmatic categories, requests, apologies, refusals, and conversational implicature, was analyzed in AntConc 3.5.9 through Word List, Concordance, N-Gram, and Keyness analysis, interpreted through Grice's Cooperative Principle, Speech Act Theory, and Relevance Theory.

Findings identify a recurring tripartite AI Limitation Formula (ALF) — softener, AI-identity disclaimer, limitation statement — that delays and obscures illocutionary force, with “as an AI language model” the dominant N-gram cluster (f=37) and however the most frequent marker overall (f=64). Keyness analysis ranks implicature as the most inferentially demanding category, since explicit markers are entirely absent, followed by refusal, whose force is most systematically concealed through the ALF and however-pivot (model: LL=45.85). Apology language proved highly formulaic (apologize: LL=23.36, %DIFF=+1689%), frequently failing Searle's sincerity condition. The study concludes AI-generated discourse encodes meaning through systematic, learnable structures with direct implications for EFL pragmatic instruction.