AI, Humor, and Metaphors | Why Can't Machines Get the Joke?

The semantic gap refers to the difference between an AI's computational processing of data and genuine human understanding of its meaning. AI, particularly large language models, excels at recognizing statistical patterns in vast datasets. It learns the associations between words, allowing it to generate grammatically correct and contextually relevant sentences. However, this is a form of sophisticated pattern matching, not a deep comprehension of concepts. For instance, an AI can process millions of texts about "love" and describe it accurately, but it does not experience the emotion or understand its profound social and biological significance. Humor and metaphors are deeply reliant on this layer of meaning that AI cannot access. They require an understanding of unspoken social rules, cultural context, emotional states, and shared human experiences. AI lacks this repository of 'common sense' and lived experience, leading to a gap where it processes the words but misses the intended meaning, especially when the meaning is non-literal. This limitation is not a matter of insufficient data but a fundamental architectural difference between artificial neural networks and the biological brain.

The human brain employs a sophisticated, bilateral system to interpret ambiguous language. The left hemisphere, particularly areas like Wernicke's and Broca's areas, is primarily responsible for processing literal, grammatical, and syntactical aspects of language. It deciphers the dictionary definition of words and sentences. However, when faced with non-literal language like a metaphor or a joke, the right hemisphere becomes critically active. The right hemisphere excels in understanding context, prosody (the rhythm and intonation of speech), and abstract connections. It helps us see the relationship between seemingly unrelated concepts in a metaphor. The prefrontal cortex (PFC) acts as an executive controller, integrating information from both hemispheres, evaluating social context, and inhibiting a purely literal interpretation to arrive at the intended, nuanced meaning. This dynamic interplay allows the brain to remain flexible, moving from literal to figurative understanding seamlessly—a capability that current AI architectures do not possess.

Humor comprehension is not a single event but a multi-stage cognitive process, often explained by the incongruity-resolution theory. First, the brain must process the initial setup of a joke, holding it in working memory and forming an expectation. Second, the punchline introduces information that is incongruous, or conflicts, with the established expectation. This violation of prediction triggers a sense of surprise, which is detected by brain regions like the temporoparietal junction (TPJ). Finally, the frontal lobe, particularly the prefrontal cortex, engages in a cognitive shift to resolve this incongruity. It finds a new, unexpected logical connection that makes the surprising punchline make sense in a playful context. This entire process requires semantic knowledge, cognitive flexibility, and 'theory of mind'—the ability to understand the joke-teller's intention.

Neuroimaging studies reveal that a distinct network of brain regions collaborates to process humor. The detection of incongruity activates cognitive regions, including the posterior temporal lobe and the inferior frontal gyrus. Following this, the resolution phase engages the ventromedial prefrontal cortex (vmPFC), which is key to creating a coherent new meaning and appreciating the joke. Crucially, the vmPFC is also connected to the brain's reward system, including the nucleus accumbens. When the incongruity is successfully resolved, this reward circuit releases dopamine, generating the feeling of mirth and pleasure associated with laughter. AI systems lack this integrated cognitive-affective loop; they can be trained to recognize the structure of a joke, but they do not experience the rewarding emotional payoff that defines genuine humor appreciation.

Teaching an AI to understand metaphors in a human-like way faces a significant hurdle known as the 'embodiment problem.' Human understanding of abstract concepts is often grounded in physical, sensory, and motor experiences. This is the core idea of embodied cognition. For example, we understand the metaphor "she has a warm personality" because we have a direct physical experience of thermal warmth, which we associate with comfort, safety, and closeness. Our conceptual system maps this physical sensation onto an abstract social quality. Similarly, a "heavy topic" is understood through our experience with physical weight. Since AI models are disembodied—they do not possess a body, senses, or the ability to interact with the physical world—their knowledge is purely abstract and derived from statistical correlations in text. They can learn that "warm" is often associated with positive personality traits, but they cannot ground this knowledge in the rich, multi-sensory experience that gives the metaphor its meaning for humans. Until AI can develop a form of embodied, sensory-based learning, its understanding of metaphors will remain superficial and fundamentally different from our own.