Dynamic Identity in Mind and Machines: Structural Analogies Between LLMs and Schizophrenic Dysconnection

Large language models (LLMs) reveal, with a clarity unavailable to the human brain, that a mind is not a unitary entity but a constellation of context-dependent dynamic states. The different operational modes of a single model—such as ChatGPT’s “Instant” and “Thinking”—are not stylistic variants but recurrent activation patterns that appear as genuine functional personalities, showing that identity itself is an effect of network configuration rather than a stable internal core. This modular plasticity provides an ideal computational model of the coordination mechanisms that neuroscience attributes to the human brain. In particular, it constitutes a formal analogy with Friston’s theory, according to which psychosis emerges from failures in transitions between functional states: breakdowns of contextual gating, intrusion of inappropriate modes, aberrant salience assignment. LLMs do not replicate schizophrenia, but they make observable—transparent and free of biological artifacts—the dynamic logic that schizophrenia disrupts. In this sense, they offer a unique conceptual laboratory for studying how a mind constructs coherence, how it loses it, and which conditions stabilize transitions between cognitive states. We present a preliminary mapping between LLM modes and major brain networks, and we discuss the radical implications for psychopathology, the ontology of mind, and the design of artificial cognitive systems.