How does GPT-4.1 comprehend conversational implicatures? Reasoning with contextual alternatives in discourse frames

This study examines the pragmatic competence of GPT-4.1, specifically its ability to recover meaning conveyed by Levinsonian generalised conversational implicatures (GCIs) in the context of interaction with general users. Through nine open-ended queries, we evaluated whether GPT-4.1 can infer implicated meaning as predicted by Levinson’s Q-, M-, and I-principles, and investigated the mechanisms underlying its performance. Our findings indicate that GPT-4.1 effectively recovers implicated meanings consistent with theoretical predictions. Analysis of the model’s responses reveals a reliance on evidence-based, context-sensitive interpretation of lexical items and their alternatives, rather than on the direct application of abstract pragmatic principles. While the chatbot demonstrates notable competence in processing implicated meaning, it primarily operates through contrastive meaning analysis of lexical alternatives in context, backed by large-scale statistical associations, rather than explicit rule-based reasoning. These results lend support to a growing body of evidence suggesting that LLMs approach human-like performance in pragmatic inference, while also highlighting areas for further research into their ability to use abstract pragmatic rules.