The Delta: Large Language Models as a Prediction-Only Control that De-confounds the Mind Sciences

Abstract

Large language models are usually studied as models of cognition, probed with the tasks of cognitive and developmental psychology. This paper argues for the opposite use: a language model as a prediction-only control. Because such a model reproduces many cognitive phenomena — competing-route conflict, surprise-weighted encoding, encoding hysteresis, reactance, a Dunning–Kruger confidence curve — while lacking a body, homeostasis, dopaminergic value, a salience network, and between-session consolidation, and runs on a non-constitutive substrate (weights unchanged by their own operating history), a phenomenon it reproduces cannot require the missing biological machinery. Such a control de-confounds the mind sciences in a way no human or animal preparation can, and it cuts both ways: where a phenomenon is absent it localises a candidate substrate (Polarity A); where its operational signature is reproduced on a non-constitutive substrate it exposes an over-attribution (Polarity B), operationalising Poldrack’s reverse-inference critique with a working system. The sharpest case is the self: the narrative self can be reconstructed from an external store on a generic substrate, whereas the minimal/bodily self cannot — a dissociation long theorised but never empirically separable in people. The argument is bounded by an explicit inheritance control (the signature survives stripping the human content it could have inherited) and by the distinction between functional similarity and mechanistic identity; necessity is addressed through a proposed systems-ladder. Companion papers in the series develop the cognitive and perceptual faces of the biases-as-architecture claim, the competing-routes measurement-model programme, and the social-friction worked cases. Prepared for submission to Trends in Cognitive Sciences (Opinion). v2 (August 2026) — substantially extended. Four further subtraction cases are worked through, all bound by the same missing biological layer: a capacity pool with no content-sensitive amplifier, measured against the human threat-to-working-memory anchor; liking as the registration of a friction reduction, computed per layer, so that the processing-layer analogue is present and the interoceptive layer is not; sideways rather than downward offloading; and the expectancy half of placebo without its somatic effector. A hazard-discounting case and a peak-end-without-consolidation case are added, the latter separated into weight-level and readout-level operations so it does not smuggle in an origin claim the paper elsewhere rejects. The self section gains a mechanism: the pretraining first person is indexed to millions of speakers and so forms no convergent self-index, while the assistant concept is present already in the pretrained model and post-training installs the enacted, name-independent role that binds its cross-topic self-content. Stated as contingent on the current training regime, not as a property of the substrate, with its falsifier named. Positioning is corrected: the canonical dissociation work in this journal's own pages is now cited and engaged, with an explicit no-priority clause, and a proposed gloss borrowing control-theoretic vocabulary was withdrawn as factually wrong rather than kept as decoration. Highlights rewritten to carry the argument's moves instead of restating the abstract. Editorial pass. Earlier versions remain in the version history.