Conceptual blending in humans and language models

In theories of embodied cognition, concept integration, or blending, integrates features from different source concepts to construct a new, context-sensitive meaning and is considered a core mechanism of human conceptual flexibility. Blending, while a cognitive act, is reflected in language, including that produced by large language models (LLMs). The seeming naturalness of blends produced by language models raises a question about underlying mechanisms of blending in brains versus models, explored in this conceptual analysis. For humans, blending is related to the embodied nature of cognition: we are constrained to perceive the environment in terms of affordances that may achieve our goals. Blending combines perceived affordances into a new concept that can satisfy the goal. I develop a predictive processing account of human generative-causal blends in which goals are parent nodes in a generative hierarchy, predicting affordances that influence perception and categorization. The account is illustrated with a toy example implemented as a vector symbolic architecture in a Python script. In contrast, language models are optimized for next-token prediction and produce distributionally novel blends constrained by proxy-goals that are derived from alignment and prompt framing within the context window. I discuss distinctions bearing on conceptual flexibility, agent autonomy, and their implications for agent design.