Semantic categorisation – the grouping of entities by meaning – is a foundational cognitive process for both biological and artificial intelligences. This essay surveys how categorisation operates in non-human systems: from its evolutionary origins and neural basis in animals, through its treatment in symbolic and connectionist AI, to its emergence in modern deep learning, particularly transformer-based language models. We define semantic categorisation and discuss why it matters for understanding cognition and AI. In Animal Models, we review evidence that vertebrates such as birds and primates form conceptual categories, describing behavioral experiments (e.g. pigeons sorting paintings or monkeys differentiating morphed images) and neural correlates of categorisation (e.g. category-selective prefrontal neurons). We note that some animals can use rudimentary symbols (e.g. lexigrams) but remain constrained by perceptual input. In Artificial Intelligence, we contrast symbolic AI, where categories are manually defined (e.g. ontologies, frames) and support explicit reasoning, with connectionist deep learning, where categories emerge from training data as distributed patterns. We then focus on large language models (LLMs): explaining how transformer architectures (token embeddings, multi-head attention, deep layers) yield rich semantic representations. We review theoretical debates about whether LLMs truly “understand” concepts (e.g. Searlean challenges, Bender Koller’s “Climbing towards NLU” vs. evidence that LLMs’ internal spaces align with human concepts (Human-like conceptual representations emerge from language prediction)), and empirical studies showing human-like categorisation effects (basic-level preference, typicality, fan effect) in LLM outputs (Large Language Model Recall Uncertainty is Modulated by the Fan Effect). In the Comparative Discussion we highlight parallels and contrasts among humans, animals and AI: all systems perform generalisation over categories, but humans use language and abstract reasoning, animals rely more on perception and evolutionarily shaped categories, and AIs rely on statistical learning. Finally, we consider implications for philosophy of mind and cognitive science, suggesting that a “piecemeal” approach (Lee, 2023) focusing on specific capabilities (like categorisation) may bridge biological and artificial models of cognition (What is cognitive about ‘plant cognition’? | Biology Philosophy ).
J. Shirley David (Mon,) studied this question.