Brain metaphors generated by humans and large language models: Evidence from metaphor elicitation
TL;DR
This study compares metaphorical descriptions of the brain elicited from 61 human participants and eight publicly available LLM-based chat systems and indicates both the capacities and the constraints of LLMs in metaphorical “reasoning,” particularly their reliance on conventionalised patterns and their limited embodied metaphorical grounding.
Abstract
Metaphor production constitutes a useful analytical framework for examining how humans and large language models (LLMs) structure meaning beyond literal description. This study compares metaphorical descriptions of the brain elicited from 61 human participants and eight publicly available LLM-based chat systems (40 AI-generated responses). Using matched elicitation prompts (e.g.,“The brain is like…”), metaphorical lexis was identified following the MIPVU procedure, and responses were categorised according to their underlying source domains. Both datasets demonstrate a marked tendency toward entrenched technological metaphor patterns, most notably the brain is a computer metaphor. However, human responses draw on a broader range of everyday and embodied source domains (e.g., brain as a muscle, sponge, or knitting yarn), whereas LLM outputs predominantly employ more abstract, system-oriented analogies (e.g., brain as a network, city, control centre) and more frequently render cross-domain mappings explicit. AI-generated responses are also substantially longer, averaging approximately five times the length of human responses. Under these elicitation conditions, LLMs reproduce conventional explanatory metaphors but exhibit reduced diversity in source domain variation and fewer mappings grounded in experiential domains. The results indicate both the capacities and the constraints of LLMs in metaphorical “reasoning,” particularly their reliance on conventionalised patterns and their limited embodied metaphorical grounding.