Linguistic differences in humour: a feature-based comparison between human and AI-generated jokes

Abstract

Humour is one of the most creative and socially embedded forms of human communication. With the rise of large language models (LLMs), questions about whether artificial intelligence can identify and generate jokes have become central to computational humour studies. This paper conducts a systematic comparison of humanand AI-generated jokes across multiple linguistic dimensions: lexical, syntactic, affective, semantic, and prosodic. To this end, we first constructed a parallel joke corpus by instructing LLMs to extract meta-information from human-authored texts and generate comparable jokes. By applying Elastic Net logistic regression on our parallel corpus of human and AI humour, we found that AI-generated jokes relied on connectives and retrospective narration. In contrast, human humour was characterised by greater lexical density, conditional subordination, emotive language, and semantic incongruity, as well as an increased use of nouns, interrogative words, social references, and intentional rhythmic devices. These findings highlight the gap between human and AI humour in terms of creativity, emotional depth, and prosodic design. The study contributes both to the theoretical validation of humour mechanisms and to the practical advancement of explainable and human-like humour generation in LLMs.