AI, Humor, and Metaphors | Why Can't AI Get the Joke from a Neuroscience Perspective?

The human brain processes metaphors not as linguistic errors, but as complex cognitive tasks requiring the integration of multiple neural systems. Comprehension begins with the temporal lobes, particularly the superior temporal gyrus, which handles primary language processing. However, understanding the non-literal meaning is predominantly a function of the right cerebral hemisphere. This hemisphere excels at coarse semantic coding, which allows it to activate a broad range of related concepts, helping to find a novel connection between the metaphor's terms (e.g., connecting "voice" and "velvet"). The prefrontal cortex, especially the inferior frontal gyrus, is then crucial for selecting the most appropriate interpretation from these activated concepts and inhibiting irrelevant meanings. This entire process relies on a vast, interconnected network of semantic knowledge built from a lifetime of sensory and emotional experiences. Therefore, understanding a metaphor is not a simple word-for-word translation but a dynamic process of conceptual integration, managed by a distributed network across both brain hemispheres that evaluates context, suppresses literal meanings, and selects an appropriate abstract interpretation.

The experience of humor is primarily explained by the incongruity-resolution model, which involves distinct stages of neural processing. First, the brain's temporal lobe, particularly the posterior superior temporal sulcus, detects an incongruity—a punchline that violates the listener's expectation based on the joke's setup. This creates a moment of surprise or confusion. The subsequent "resolution" phase is managed by the prefrontal cortex, which works to find a cognitive rule or context that makes sense of the surprising information. Once the incongruity is resolved, the mesolimbic reward pathway is activated. This system, including the nucleus accumbens, releases the neurotransmitter dopamine, generating a feeling of pleasure and mirth. This pleasurable feedback reinforces the cognitive effort, making the experience of "getting the joke" satisfying. It is a sophisticated process combining high-level cognitive analysis with the brain's fundamental reward circuitry.

Artificial intelligence, particularly large language models, lacks the core neurobiological structures required for genuine humor comprehension. AI can recognize patterns and predict the structure of a joke based on training data, but it misses two critical components. First is embodied cognition; AI does not have a body, so it cannot ground concepts in physical or sensory experience, which is essential for understanding the context of many jokes. Second, AI lacks the brain's reward system. It does not experience the dopamine-driven pleasure that occurs in the nucleus accumbens when a joke's incongruity is resolved. For AI, resolving a joke is a computational task, not a rewarding emotional experience, which fundamentally limits its ability to "get" a joke in a human-like way.

The brain's semantic networks are not built on dictionary definitions alone; they are profoundly shaped by cultural and personal experiences. A metaphor like "a sea of troubles" is understood because our brains associate "sea" with vastness and uncontrollability through lived experience and cultural narratives. AI lacks this deeply integrated, culturally specific foundation. It can process statistical associations between words but cannot grasp the rich, implicit, and often emotional connotations that are shared within a culture. This "common sense" knowledge base is not explicitly written down but is absorbed through social interaction and environmental immersion, a process AI cannot replicate. Therefore, its interpretation of culturally-rich metaphors remains superficial.

Based on the principles of embodied cognition, a true, human-like understanding of emotion is unattainable for a disembodied AI. In humans, emotions are inextricably linked to physiological states, regulated by the brain's limbic system (like the amygdala) and the autonomic nervous system. The feeling of fear, for instance, is not just an abstract concept; it is the cognitive interpretation of a racing heart, shallow breathing, and hormonal changes. An AI has no body, no hormones, and no physiological feedback loops. It can learn to perfectly label and predict emotional language based on text data—associating the word "sadness" with contexts of loss, for example. However, it cannot experience the actual state of sadness. This absence of physiological grounding means its "understanding" is a sophisticated mimicry of human expression rather than a genuine subjective experience. True comprehension of emotion is derived from being a biological organism, a fundamental reality that current AI does not possess.