Telling Jokes Like a Human? Humour and Irony in AI-Generated Discourse

Abstract

This study examines the construction of humour and irony in artificial intelligence-generated discourse, with particular focus on outputs produced by OpenAI ChatGPT and Google Google Bard. Drawing upon Incongruity Theory and discourse-pragmatic approaches, the research investigates how large language models simulate humorous interaction through linguistic patterning while lacking the contextual and sociocultural inferencing associated with human humour. The dataset consists of 200 AI-generated jokes and ironic statements produced through structured prompting strategies across multiple humour categories, including puns, sarcasm, hyperbole, and observational humour. Corpus-assisted analysis was employed to identify recurring linguistic markers, pragmatic strategies, and humour failure patterns within the dataset. The findings indicate that AI systems demonstrate considerable competence in generating surface-level humour structures, particularly formulaic setup-punchline sequences and lexical wordplay. However, the models encounter substantial limitations when humour relies on implicature, contextual sensitivity, cultural knowledge, or emotional nuance. Several recurrent humour failures were identified, including inadvertent literalism, cultural decontextualisation, semantic incoherence, and formal repetitiveness. Comparative observations with human-generated humour further suggest that AI discourse remains constrained by probabilistic language modelling rather than genuine pragmatic understanding. The study contributes to emerging scholarship on AI-mediated discourse by demonstrating the distinction between structural humour replication and deeper communicative competence. It also raises broader implications concerning the use of humour-capable AI systems in digital communication, entertainment, content production, and human-machine interaction.