Exploring image–text combinations in visual humour through large language models (LLMs)

Abstract This article aims to advance the scholarly discourse on the linguistic capabilities of Artificial Intelligence (AI) by analysing how three different Large Language models (LLMs) approach visual humour. Specifically, it focuses on meme generation and interpretation, with an emphasis on image–text combinations. In the first section, three LLMs were prompted to generate accompanying text for ten preselected images. In the second section, they were prompted to analyse one hundred human-made memes. Results show that, despite a generally positive evaluation in humour detection, applying non-literal meaning remains a challenge, even for the best-performing AI models.