The Internal Dynamics of Decisions: From Rats and Primates to AI

<jats:p>Whether the signatures of conscious access are fixed by what a system computes, or also by the substrate that implements it, is usually argued conceptually: empirical purchase requires the same computation examined in two different substrates, with the same signature sought in both. The unfolding argument sharpens the difficulty, since for any recurrent network there is a feedforward network with the same behaviour, so behaviour alone cannot settle claims resting on internal dynamics, which must be tested directly. A dynamics demix separating input-driven change from change generated by a population’s own recurrence — validated on synthetic systems before any recorded data — was applied to two substrates on matched cognitive work: cortical populations in perceptual and value-based decisions, and language models constructing multi-step answers. Every neural dataset was autonomous-dominated; five language models were input-dominated, a categorical regime difference. At commitment to a perceptual decision the cortical autonomous component strengthened, locked to commitment, not the stimulus; the models showed no such attractor and committed only at their output layer. In a pre-registered control varying only architecture at matched accuracy, the eigenvalue measure separated recurrent from feedforward networks — architecture, not task. Pre-registered negative tests show the cortical effect is invisible to trial-averaged and scalar read-outs. A dynamical property used to reach a commitment is present in cortex but absent from a feedforward model on an analogous task. Nothing is claimed about phenomenal experience or machine sentience; the gap between decision commitment and conscious access is the central limitation.</jats:p>