Contemporary debates surrounding large language models (LLMs) oscillate between two inadequate extremes: reductive mechanistic accounts that deny meaningful semantic organization, and anthropomorphic accounts that prematurely attribute human-like consciousness to artificial systems. This paper argues that both frameworks fail to characterize the emergent semantic behaviors observed in transformer-based architectures. We propose Operational Semantic Understanding (OSU), defined as the capacity of an artificial system to manipulate semantic content coherently, contextually, adaptively, and inferentially without requiring evidence of phenomenological consciousness. We further introduce Artificial Semantogenesis, referring to the emergence of functional semantic organization through large-scale relational learning within distributed inferential architectures. Both concepts are positioned relative to established philosophical accounts, including Dennett's intentional stance, Dretske's informational semantics, and Harnad's symbol grounding problem. We argue that contemporary LLMs may constitute a novel category of inferential artificial cognition, and propose an expanded cognitive taxonomy distinguishing Biological Phenomenological Cognition, Symbolic Computational Cognition, Inferential Artificial Cognition, and Hybrid Relational Cognition — the latter describing emergent knowledge co-production through sustained human–AI collaboration. Future research, we contend, should move beyond binary debates about whether AI systems "truly understand" and develop more precise frameworks for describing operational semantic functionality across diverse cognitive architectures.
Roca et al. (Mon,) studied this question.