Literal Words, Hidden Worlds: Implicature and Context in Generative AI Communication

Abstract

Generative artificial intelligence (GenAI) is increasingly used in education, professional communication, and everyday life. Although these systems can produce fluent responses, fluency does not always show that they have understood what a speaker intends. This paper examines conversational implicature: meaning that is suggested rather than directly expressed. Guided by Grice’s (1975) theory of conversation, the study focuses on how a generative AI system distinguishes literal meaning from implied meaning and how its interpretations may change when linguistic, situational, and social context changes. A qualitative comparative design is proposed. Approximately 20 context-dependent utterances representing indirect requests, scalar implicatures, polite refusals, and implied criticism are to be presented to one selected language model under three contextual conditions, producing about 60 responses. The responses are to be examined using four categories: accurate literal interpretation, contextually supported implicature, ambiguous interpretation, and unsupported inference. Previous studies indicate that language models can make some pragmatic inferences, but performance differs across models, phenomena, contexts, and evaluation methods (Cho & Kim, 2024; Cong, 2024; Ma et al., 2025; Wagner et al., 2025). The paper addresses the need for a focused comparison that examines literal and implied meaning together while context is varied systematically. Since the actual response dataset is not included in this document, the discussion and outcomes provide an analytical framework rather than claims of completed experimental findings. The study may help researchers and educators evaluate AI-generated interpretations more carefully and recognise when additional context is needed.