From silicon to synapse: Why the artificial unit is not a kind of neuron.

Abstract

The computational units of large language models (LLMs)-artificial neurons, attention heads, and weight matrices-differ from biological neurons not merely in degree but in kind. This paper develops that empirical argument across three levels of biological organization: the single neuron, the local circuit, and whole-brain dynamics. At the single-neuron level, biological neurons are self-maintaining electrochemical processes sustained by continuous ATP-dependent ion pumping, activity-dependent morphological remodeling, and spike-timing-dependent plasticity (STDP) grounded in local calcium dynamics at individual synapses; artificial units are stateless, memoryless scalar transformations with no metabolic basis. At the circuit level, cortical columns implement laminar excitatory-inhibitory balance and self-organized collective oscillatory dynamics that are absent from the feedforward, batch-normalized layer stacks of transformer architectures. At the systems level, the brain sustains always-on intrinsic dynamics, homeostatic regulation, and circadian entrainment; LLMs are stateless between inferences and possess no equivalent of resting-state activity. Across all three levels, the differences are organizational rather than scalar: they concern the mode of existence of the processing units, not merely the quantity or speed of processing. We conclude that the term "artificial neuron" is a historical metaphor that has outlived its usefulness as a scientific description, and that the conflation of artificial units with biological neurons is a source of persistent confusion in both AI research and public understanding. Replacing the metaphor with precise empirical description is a precondition for accurate evaluation of what LLMs can and cannot do.