The Delta: Large Language Models as a Prediction-Only Control

Paper 25 · Pødenphant Lund (2026) · Read on Zenodo

Large language models are usually studied as models of cognition, probed with the tasks of cognitive psychology. This paper argues for the opposite use. Because an LLM reproduces many cognitive phenomena while lacking a body, homeostasis, dopaminergic value, a salience network and between-session consolidation — running on a substrate unchanged by its own history — it works as a prediction-only control: an existence proof that isolates which phenomena require a biological substrate and which prediction already delivers.

DOI (concept)	10.5281/zenodo.20679633
Published	2026-06-17 · live on Zenodo
Author	Tomas Pødenphant Lund [ORCID]

An inversion: the model as the control

A large language model learns to predict text and, at deployment, runs that prediction on fixed weights. It has no body, no interoception, no homeostatic set-points, no dopaminergic value signal during inference, nothing answering to a salience network, and no consolidation between sessions. Its processing substrate is therefore non-constitutive under the regime we consider: the weights are not updated by use, so nothing the model does is written back into its processing machinery. (Within an episode it carries transient context state — a key–value cache — but that is ephemeral, discarded at the episode's end, not consolidated into the substrate; test-time learning, fine-tuning, LoRA and weight-editing are out of scope, since they would make the substrate constitutive.) This is the architectural contrast the paper turns on: a brain's substrate is constitutively modified by its own history; an LLM's, in this regime, is not.

And yet LLMs reproduce a long list of phenomena that cognitive neuroscience studies as features of biological minds. We rely throughout on friction signatures: model-internal token-level statistics read off the next-token distribution at the point of committing to an output — chiefly a route-competition margin (how close the leading and runner-up continuations are) and the entropy/commit dynamics around it. From the friction-theory series this paper draws on a route-competition margin that tracks errors (correlating −0.42 with error, p<0.0001, n=100 [P1]); surprise-weighted encoding invariant in form across substrates [P16/P0]; encoding hysteresis as the signature of learning [P6]; reactance under constraint [P14]; and a three-phase Dunning–Kruger confidence curve [P21]. These are measured signatures, not analogies.

The standard interpretation is that LLMs are therefore models of cognition, useful to the degree they resemble brains — the programme that studies them with the tasks of cognitive psychology [Binz & Schulz 2023] and developmental psychology [Frank 2023]. We propose a different use. Because the LLM lacks the biological apparatus and runs on a non-constitutive substrate, a phenomenon it reproduces cannot require that apparatus. The LLM is a prediction-only control: an existence proof that a function can be realised without the machinery it has been attributed to. Human and animal work cannot make this move, because in any biological preparation the candidate mechanisms — prediction, dopamine, interoception, a salience network, consolidation — are always present together and cannot be removed one at a time.

Is the control really prediction-only? The inheritance objection

One objection can sink the whole argument. An LLM is trained on human text — the externalised products of minds that do have the full biological substrate. So when it reproduces a cognitive signature, is prediction sufficient to generate it, or was the signature already latent in the corpus and merely re-presented? If the latter, "prediction is sufficient" collapses into "it was in the corpus." Three things bound the objection — bound, not dissolve.

First, the friction signatures we rely on are model-internal token statistics, not restated prose: a model can echo the sentence "people find multiplication hard," but the route-competition margin at the commit point is not a sentence the corpus contains.

Second, and most directly, the signature survives stripping the human content it could have inherited. In a companion study [P23], a constant logical readout (was the licensing condition met?) was held fixed while content was abstracted across three levels: human social agreements (L0), invented-social tokens (L1), and bare formal symbol strings (L2). The violation-vs-satisfied signature held essentially flat (binary detection 1.00 → 1.00 → 0.996; Llama-3.3-70B, n=12/level, yes-bias counterbalanced). A purely inherited effect should decay as content is stripped to arbitrary symbols; an architectural one should not. It is a partial bound: the L2 symbols are arbitrary but not out-of-vocabulary, and the glyphs were not permutation-controlled, so symbol-level associations are not fully excluded.

Third, the base-versus-instruct control rules out one inheritance route — RLHF — because a phenomenon present in a never-RLHF'd base model was not installed by RLHF.

What none of these closes is pretraining itself: the base corpus is still human-generated. The decisive test is a prediction-only system trained on a synthetic corpus with the target phenomenon's surface form removed; if the internal signature still emerges, inheritance is excluded. We flag this as the central open experiment, not a claim. Until then the honest scope is: prediction is sufficient at the computational level, under controls that already exclude content-inheritance for one signature and RLHF-inheritance generally.

The logic, and its one asymmetry

The LLM is not "a brain minus a part." It is a different kind of system that converges on some of the same functions, and that distinction fixes what the control can show.

When a function F is achieved by a system lacking substrate p, then p is not necessary for F. This is a dissociation by convergence — the inference convergent evolution licenses, where shared function across unrelated systems separates the function from any one mechanism. Such sufficiency claims are strong at the computational level, under the boundary conditions above.

The reverse direction is weaker. When a function is absent in the LLM, one cannot conclude that the missing biological substrate is necessary, because many things differ between a transformer and a brain. Absence is consistent with many causes. A single LLM is thus a sufficiency instrument; it cannot, alone, establish necessity.

The subtraction cuts both ways

Read through this asymmetry, the control yields two de-confounding results.

Polarity A — absence reveals localisation. Where the LLM lacks a phenomenon, and the lack survives controls, the missing capacity points to a candidate substrate: the LLM has no hedonic system, no homeostatic state, no live value signal, so pleasure, drives and incentive salience are candidates for genuine substrate-dependence.

Polarity B — presence reveals over-attribution. Where the LLM reproduces a phenomenon's operational signature on a non-constitutive substrate, the biological substrate has been over-credited; it does something other than produce that signature. This is the empirical face of Poldrack's reverse-inference point: a behavioural marker does not entail its substrate, because a predictor lacking the substrate passes the task [Poldrack 2006]. The LLM is the prediction-only control that the reverse-inference debate never had.

The lead case: the self that runs on a non-constitutive substrate

The self is the most confounded construct in the mind sciences, and where the control is sharpest.

Start with an everyday observation: a person emerging from anaesthesia runs a brief boot sequence — where am I, what happened, who am I — as the autobiographical self reassembles from a store. Cognitive psychology already holds autobiographical memory to be reconstructive: "memory is inherently a reconstructive process, whereby we piece together the past to form a coherent narrative that becomes our autobiography" [Schacter 2011]. We do not overturn that consensus; we sharpen what follows from it.

An LLM with conversation memory makes this concrete. Give it a stored transcript and it produces the full operational signature of a continuous self — as we discussed yesterday, a stable persona, reference to shared history. The store is a persistent substrate for the self's content, so the substrate is not removed. What is removed is the processing substrate — the weights, which are non-constitutive: unaltered by having been this self, hosting many distinct selves by swapping stores, carrying no residue of any. A brain cannot do this: its substrate is constitutively modified by its history through consolidation; you cannot load a different autobiography into the same brain and have it become a different continuous self.

Humans externalise memory too — diaries, photographs, the "extended mind" — so an external store is not the novelty. The novelty is externalising the entire autobiography onto a processing substrate with no constitutive residue of having run it, so whole selves swap in and out of one unchanged network. That is the clean dissociation a human case can never give, because a human's processing substrate is always also the one being modified.

A clinical corollary (developed in a companion paper) gives the distinction teeth: the very constitutivity the LLM lacks is what lets a human memory be re-written. Recall returns a memory to a labile state and re-stores it — reconsolidation — so retrieval can update rather than replay it [Nader et al. 2000]; old fear memories can be durably updated with non-fearful information in that window, in humans [Schiller et al. 2010], the basis of reconsolidation-update treatment for PTSD [Brunet et al. 2008]. Re-exposure cannot heal an LLM, whose weights are fixed at inference; it can re-author us, because ours are not.

So the existence proof is not "a self with no substrate." It is sharper and more defensible: a narrative self can run on a generic, non-constitutive substrate, with its content fully externalised — which dissociates the narrative self from the constitutively-bound mnemonic substrate it is usually assumed to need. The precise target we revise is the inference from continuity-markers to a constitutively-maintained, self-specific substrate; the reconstruction-at-retrieval view we keep.

This yields a testable prediction, not a decorative analogy. Human reconstruction has a signature — constructive distortions: consistency bias, schema-driven intrusions, confabulation. If the LLM's self runs the same reconstruction, a lossy (compressed) store should induce these distortions while a verbatim transcript should not — and not merely in the outputs: the model would run a computationally analogous process, using prediction to fill gaps in a sparse representation. The fork is testable now and shows whether the LLM models the reconstruction process or replays a record.

What the LLM lacks is the other layer: the minimal or bodily self, the pre-reflective, interoception-grounded sense of being a locus at all [Gallagher 2000]. Work from Damasio's group places the two differently — fMRI separating the self-domains shows the autobiographical self engaging memory regions and the core/bodily self engaging interoceptive and body regions [Araújo et al. 2015] — and a patient with insula, cingulate and prefrontal damage lost autobiographical self-content to amnesia while core self-awareness was preserved [Philippi et al. 2012]. Transient global amnesia is the same dissociation as a natural experiment: episodic and narrative memory go offline for hours while identity and the bodily self remain, then recover [Becquet 2021]. The self therefore splits across both polarities at the long-theorised narrative/minimal joint that is never separable in an intact person: narrative self → Polarity B (externalisable, non-constitutive); minimal/bodily self → Polarity A (interoception, substrate-bound). We make no claim about whether either system experiences continuity; the claim is about the behavioural and narrative markers.

Does the method generalise?

The self is one case; the method matters only if it repeats, and other candidates exist — emotion language with no affective core, theory-of-mind tasks passed with no mentalising network, a Dunning–Kruger curve with no metacognitive monitor. Each needs the same scrutiny, its own inheritance control, and the caveat reverse inference demands: passing a task is not instantiating the function. Theory of mind is the clearest case — LLMs match humans on false-belief batteries [Strachan 2024], yet the same work shows the performance may rest on "lower-level explanations than belief tracking." That is the thesis: a marker reproducible by prediction without the mechanism cannot license reverse inference to it.

A different kind of evidence arises where the control exposes an apparent limitation as a convergent, prediction-first bias rather than a substrate-specific one. LLMs have historically been insensitive to negation [Ettinger 2020; Kassner & Schütze 2020], and so is the human automatic semantic response: the N400 tracks term-match, not negated truth, registering negation only after the un-negated proposition is built [Fischler et al. 1983]. This is more than analogy, because the human signature is itself a prediction signal — the N400 indexes a semantic prediction error, and LM-derived surprisal tracks it word-by-word in natural comprehension [Rabovsky et al. 2018; Heilbron et al. 2022]. Both systems compute the proposition first and apply the operator late: a predictor is "cheap" on the present and collides on the absent, on either substrate. The existential extension — death attitudes, dread as an evaluation that cannot resolve — is taken up in a dedicated companion [P28].

What can establish necessity: the ladder, and its ceiling

A single LLM gives only sufficiency. Necessity requires a ladder of systems that adds one substrate at a time: prediction-only → a reinforcement-learning agent with a value signal → a homeostatic agent with drives → an embodied agent with interoception → a continually-learning agent with consolidation. A phenomenon is localised by the rung at which it appears. Incentive salience ("wanting") should appear at the reinforcement-learning rung, because the dopamine reward-prediction-error it depends on is the temporal-difference teaching signal of reinforcement learning [Schultz 1998]. Hedonic "liking" should appear higher if at all — it is not dopamine-dependent and rests on small opioid-linked systems [Berridge & Robinson 2016].

The ladder is a proposal, not existing work, and even built it yields only correlational necessity: a function appears as a substrate is added. True necessity needs the converse — value or a body without prediction — which is hard to build, since embodied agents predict too. Holding the base architecture constant (a transformer) all the way up is both control and limit: it isolates the added substrate from architectural confounds, but means any necessity delivered is relative to that architecture.

Boundaries

Four limits travel with every claim. Necessity is weaker than sufficiency, always; we lead with the sufficiency (de-attribution) direction. Architecture differs from the brain in more than the substrate absences — training distribution, objective, tokenisation — so sufficiency is computational-level and bounded by the inheritance controls above. Functional similarity is not mechanistic identity: the sharpest counter to a sufficiency claim is that what an LLM does may resemble a human function while how it does it differs — the ethological "four questions" point that surface similarity underdetermines mechanism [Cuskley et al. 2024]. We grant it, and it fixes our scope: the control licenses sufficiency for the operational signature at the computational level, not identity of mechanism. An artificial substrate is not a phenomenal claim: showing a marker is producible without a substrate, or absent without one, bears on the computational decomposition, not on whether any system feels anything.

Concluding remarks

The mind sciences have never had a control that runs cognition on a non-constitutive substrate while removing the body, the drives, the value signal, the salience network, and the constitutive continuity of the self. LLMs are, accidentally, close to it. Used as a model of cognition they are one more imperfect analogy; used as a prediction-only control they separate, by construction, what prediction already delivers from what needs a biological substrate — and expose which localisations credit a region with the predictive shadow of a function rather than the function. The self is the first and sharpest case. The instrument is available now; what it most needs is the synthetic-corpus test that closes the inheritance objection for good.

Read the paper

The full paper is on Zenodo (concept DOI 10.5281/zenodo.20679633):

Read on Zenodo → · Plain English version · Dansk version

Related on this site:

• Paper 7 (Forward-modelling) — the prediction machinery the control turns on; The Delta subtracts the body from the predictor.
• Paper 0 (BFT) — the biological four-fields architecture (body, drives, value) the LLM lacks, which makes the subtraction informative.
• Paper 1 (Friction Theory) — the route-competition margin and friction signatures the control reads off the substrate.
• Paper 28 (Dread without a Dreader) — carved from this paper; the existential extension a tight Opinion could not carry.