Tong argues in her research that fluent language production by AI is not the same as language understanding. Metaphors and humour are not merely decorative features of communication; they help people explain complex ideas, express emotions and maintain social relationships. Expressions such as “letting something go” or “defending a position” are used daily without people consciously reflecting on their figurative meaning. Humour involves similar mental processes: recognising unexpected connections, contradictions or shared cultural references. According to Tong, this comes naturally to people but remains surprisingly difficult for AI systems.

To investigate this, Tong developed new datasets and evaluation tools. One of these is the Metaphor Understanding Challenge Dataset, known as MUNCH, which contains more than 10,000 carefully annotated paraphrases of metaphorical sentences. Tests with well-known language models showed that these systems frequently struggled to distinguish between literal and figurative meanings and often failed to grasp the intended meaning. Tong also investigated whether AI can recognise why a metaphor is being used, given that metaphors can serve a range of functions, from clarifying complex concepts to strengthening social bonds. Together with colleagues, she compiled the first large-scale taxonomy of metaphorical intentions, divided into nine categories. Current AI models could recognise some of those intentions reasonably well, but in many cases ran into clear limitations.

Humour proved a particular challenge when both text and images are involved. For this part of the research, Tong compiled the so-called Hummus dataset, a collection of 1,000 cartoons from The New Yorker. Many of those cartoons rely on visual metaphors and subtle connections between image and text to create humorous effects. Even highly advanced AI systems often struggled to combine visual and textual information in a way that allowed them to understand the joke. Tong concludes that grasping a joke requires cultural knowledge, social context and a sense of what makes a situation unexpected, all of which remain beyond the reach of current systems.

The findings carry practical relevance as AI systems take on a larger role in virtual assistants, educational tools and workplace applications. Tong concludes that future AI systems can only become truly reliable, intuitive and useful once they learn to better understand metaphors, humour and cultural context. When an AI assistant misses those subtleties, it can affect the quality and reliability of interactions with the people who use it.