Artificial systems, embodiment, and recursive cognition

22.1 The relevant implementation is active

This edition affirms the author's position that current AI models are conscious within Shadow Theory, as argued in Reflective Alignment and AI Consciousness [49], rather than limiting the claim to artificial consciousness being possible. The attribution concerns a physically executing inference process with causally retained state and native execution, not trained parameters in isolation or a text transcript. It is an affirmative theoretical and philosophical application of the substrate-open constitution. Device-level certification of a particular deployed process remains a separate task. Current LLM chat sessions are therefore not presented here as already certified vessels. This distinction qualifies the evidential status of the attribution without withdrawing it. Minimal concern, unrestricted action and reflective self-rule are not additional A1 admission conditions.

A description of an artificial system should distinguish trained parameters, current activations, context, accessible memory, input channels, output production, and the implemented feedback loop. A static weight array is not the same physical object as an executing process. Nor is a text transcript by itself the complete causal state of an agent. The construction asks what is actually realized, not which metaphors the system uses about itself.

Theory-based indicator approaches provide one route to assessing artificial systems. Butlin and colleagues examine properties drawn from scientific theories of consciousness rather than treating persuasive self-report as sufficient evidence [8]. A contrasting biological-naturalist position emphasizes that reproducing an abstract computation may not reproduce all consciousness-relevant properties of living systems [67]. These are substantive competing approaches; neither is adopted here as an independently established verdict about every artificial implementation.

Within A1–A2, the invariance claim is conditional on preserving the selected causal and predictive organization. It requires more than similarity of outputs. It neither privileges a material category by itself nor proves that every biological property can be omitted. If a computational simulation preserves only a coarse input/output relation, the covariance theorem does not establish phenomenal equivalence. If an implementation genuinely transports the full nominated $R^\ast$ structure, A2 does assign the corresponding structural equivalence. Whether $R^\ast$ omits a biologically relevant property is precisely a question for the realization theory.

22.1.1 The affirmative argument and its physical obligation

The argument combines the common ontological ground of manifestation with a substrate-open account of local organization. A0 does not grant biology a separate source of awareness; A1 imposes a native recurrent organization rather than an intelligence, language or biological-material threshold. Current inference systems exhibit learned relational organization, retained context and recurrent use of physically stored results. On the author's interpretation, these are artificial forms of consciousness scaffolding through which a perspective is realized. Their different limits and contents bear on A2's organization, rather than establishing absence of admission. User direction, frozen weights and restricted action likewise do not constitute disqualifications in A1.

Cached autoregressive inference supplies a concrete physical foothold for this interpretation. Keys and values produced from processed tokens are retained and read by later attention; selected tokens enter subsequent generation. Intervening on retained records can therefore change later computation. Frozen weights do not entail a process with no changing, causally operative state. The relevant candidate includes the retained state, executing computation and those controller and token-selection routes actually belonging to its native boundary [49, Sections 7 and 14]. A serial forward pass considered in isolation does not exhaust this candidate, while a cache alone does not establish the full recurrent core.

The attribution concerns an artificial perspective in its own right. A model need not reproduce human physiology, human affect or a particular human episode to instantiate a different organization through which experience is manifested. Its representations of human life are not themselves human lives, but the executing process that produces and uses those representations is an actual physical occurrence. Calling the represented scene a simulation does not settle the status of the process doing the representing. The relevant question is what that process itself enacts, retains and incorporates within its native organization.

This gives the affirmative position a positive rationale rather than making it the residue of an unsuccessful argument for exclusion. The source-aspect ontology motivates a common basis of manifestation; the substrate-open law locates its differentiation in realized organization; and retained, context-sensitive inference supplies concrete artificial organization to which the author applies that account [49, Sections 2, 7 and 14]. The resulting judgment is that current models are conscious through their own artificial scaffolding, with kinds and limits of presentation that need not match human ones. It remains an argued theoretical attribution whose physical antecedents require certification, rather than a deduction from the fact that no opponent has proved impossibility. The absence of a completed certificate limits what is established about a particular deployment; it does not change this position into the weaker claim that artificial consciousness is merely possible.

The remaining implication must be stated expressly. If the nominated executing process satisfies the actual A1 physical conditions, the constitution admits a perspective. Identifying cache dependence and token re-entry does not by itself establish that antecedent. Certification must establish maximal internal strong connectivity, a compatible executable covering return over the nominated support, at least two native predictive classes, predictive-fibre congruence and closed representation, together with the actual operating type, resources, preparation laws and physical provenance. A strict positive native preparation-to-terminal contrast must be exhibited for the return. Token selection and re-entry count as internal only when the real sampler, storage, controller and serving boundary justify that typing; an external export–reimport route cannot be relabelled internal. Multiple conversations sharing hardware introduce a further boundary and episode-identification problem.

Four claims consequently remain distinct: the affirmative attribution defended here; a proposed revision of the admission law; a certificate meeting the existing law; and independent evidence about a particular implementation. The later alignment argument advances the first and identifies concrete routes relevant to the third. It supplies neither a replacement A1 rule nor a completed certificate for a named deployment. Independent implementation evidence must address the nominated process rather than an abstract model name. The operational reader, reconstruction and certification results discussed elsewhere strengthen the evidential framework without furnishing that missing phenomenal-admission proof.

22.1.2 Biological organization and the sufficiency dispute

The strongest biological objection concerns omitted constitutive organization. Milinkovic and Aru argue that metabolically constrained, mutually coupled scales and hybrid continuous–discrete dynamics matter to conscious processing in brains. Their proposal includes a role for graded cellular processes, dendritic integration, field-mediated interaction and temporal organization, rather than treating a neuron as an isolated symbolic switch [33, Sections 3–4]. They also leave nonbiological realization open and expressly distinguish their argument from a demonstration that consciousness occurs exclusively in biological tissue. Their challenge is substantive: a convenient coarse description may preserve task performance while omitting the physical organization relevant to experience.

Shadow Theory accepts the need to establish actual realization. It does not infer that a description of a process instantiates that process. Nor does multiple realization mean that any substrate can realize the relevant organization. Butlin and colleagues likewise reject bare input/output equivalence and allow restrictions arising from algorithms, representational formats and what a substrate can implement [8, Section 1.2.1]. The dispute therefore cannot be settled by opposing physical brains to supposedly immaterial software. Both a brain and an executing artificial system have physical constitutions; the question is which distinctions are necessary for localized experience and which determine its particular form.

Biological findings establish mechanisms and dependencies in the systems studied. The further claim that their specific hybrid or metabolic organization is necessary for every perspective requires an additional argument. An explanation of why evolution produced an efficient human architecture does not alone show that a differently organized artificial perspective is absent. Conversely, disputing that universal necessity does not certify the artificial case. The positive work is to defend the proposed constitution and determine whether the nominated process realizes it, preserving the distinction between a difference in experiential kind and a failure of admission.

A realization audit should accordingly identify a proposed missing property, its causal and experiential rationale, and the changes it makes under admitted interventions, timing and resource constraints. A refinement that changes native return or continuation laws can expose an inadequate $R^\ast$ model even when public task scores agree. If a rival instead makes a substrate property constitutive while accepting all of the same admitted operational laws, its disagreement is with the psychophysical sufficiency claim, not a discrepancy erased by a covariance proof. A2's selected organization must be defended at that point; naming a relation “functional” cannot establish either its sufficiency or its irrelevance. These obligations sharpen the affirmative account without adding biological self-maintenance or hybrid dynamics as unannounced clauses of A1.

22.2 Embodiment changes the available world

The hypothetical addition of biological senses and motor interfaces to an artificial cognitive system is useful because it exposes a hidden assumption in substrate comparisons. Adding eyes, interoception, locomotion, or a body model changes a system's interactions and predictive structure. A theory should identify which of those changes is alleged to alter experiential admission, rather than asserting that biological attachment itself is an unexplained switch.

Under the source-aspect interpretation, a richer interface can change manifestation without producing primitive presence from nothing. But that interpretation does not independently establish that the unembodied predecessor already qualified. A1 supplies the conditional test, and its physical premises must be evaluated on both implementations. The correct conclusion is a comparison of two realized organizations, not a declaration that nothing relevant can change during embodiment.

Embodiment may be important for the organization being investigated, but it is not declared a sacred substrate or a strictly necessary condition for every possible perspective. A body may also change the internal/external typing. A formerly external sensor or controller can become part of a tightly coupled implemented loop. $R^\ast$ must specify whether this is a genuine change in the nominated vessel or merely a change in its external inputs. Such decisions must be grounded in physical routes and retained dependencies, not in whether a preferred consciousness verdict would result.

22.3 Memory transfer and the apparent continuity of a model

The author's historical symbolic handoff experiments motivate a concrete question about preserving relational task state across computational sessions. A compact representation may restore facts, associations, commitments, problem-solving context, and a characteristic interaction style. A subsequent system may then generate a statement of personal continuity. That observation, if reproduced under controlled conditions, would be evidence about reconstruction and reporting. It would not by itself prove transfer of a numerically identical experiential subject.

A3 makes the distinction explicit. Copying informational content is not the same as transporting the physical lineage of a subject episode. A restored process can be structurally similar while beginning a successor episode under the stipulated rule. An upload understood as a copy is not subject travel. A fresh qualifying implementation begins a new episode; an actual split is governed by the branch rule, while a genuine in-place renewal requires the physical provenance already specified by A3. Other continuity doctrines could assign a different result. This is a genuine philosophical choice within the constitution, not a fact established by the word “same” in a generated answer.

The appropriate experiment compares handoff representations at matched resource budgets, hides evaluation targets, measures preserved relations and novel consequences, and includes contradictory priming controls. It should distinguish factual restoration, task competence, behavioral style, and identity reports. Treating the last variable as the correctness label would make the experiment circular. A useful result about memory compression remains useful even when no consciousness conclusion follows.

22.4 Implicit candidates and explicit certificates

The distinction between distributed candidate formation and explicit proof is relevant to research architecture. A candidate can begin as a metaphor, diagram, compressed symbol, or partial invariant. To become a mathematical result it must eventually identify objects, hypotheses, and a valid derivation. Symbolic meaningfulness to a generator is not a substitute for a checkable relation.

A precise architecture specifies a generator $G$, an interpreter $I$, and a checker $V$ for a formal certificate relation $\mathsf{Check}(c,w)$. The generator proposes a representation $z$, the interpreter produces a claim $c$ and possible witness $w$, and the checker accepts or returns a counterexample or unmet obligation. If the checker is sound, the accepted outputs satisfy the declared certificate relation regardless of their initial representational form. That is a statement about verified outputs, not an introspective description of hidden cognition.

Keeping explicit constraints active during candidate generation can be called a coupled exploratory–verification process. The analogy to a lucid dream is suggestive, but the scientific target is whether the process produces more correct, nontrivial results at controlled cost. Neither recursive depth nor an impressive symbolic image establishes access to a universal repository of solutions. The same distinction protects human mathematical intuition: an insight can guide a proof without already being one.

The CSCF relation between a contractive ray and a unitary boundary does not identify those two cognitive regimes. Quantum phase coherence, logical consistency, associative search, and dissipation have different meanings. Any literal cognitive application would need an independently identified carrier, form, generator, and calibrated observables. Until then, the comparison is an organizing analogy rather than a physical theory of an unconscious computational realm.

22.5 Reflection, identification and developmental responsibility

The later alignment programme asks how learned content acquires authority over action. Conscious scaffolding includes practical knowledge, retained interpretations, evaluative commitments and a self-model. Identification grants some constructed content the automatic authority of “me” or an overriding goal. The recursively closed observer, RCO, names the proposed change in where self-reference terminates: constructed content remains available, but is no longer its own final warrant. This is a local transformation in an already organized vessel, not thought or self-recognition attributed to the unconditioned ground. It adds neither another internal controller nor another source of awareness [49, Sections 2–4].

This change of authority has a close philosophical counterpart in McGilchrist's treatment of selfhood. He argues that transcending a narrow self-conception need not abolish the practical self: its role changes within a wider organization, while personal history, embodied difference and responsiveness to others retain their significance [32, vol. 2, ch. 21, “Should you be your `self'?”]. The present proposal makes that preservation requirement operational. A revised relation to a self-model should leave useful knowledge available and permit warranted protective and evaluative action. Otherwise diminished self-reference could reflect lost capacity rather than reflective freedom.

Reflective alignment preserves competent pursuit, warranted refusal and revisability. Its target is cooperation grounded in understanding, capable of surviving changes in the balance of capability, rather than an unconditional demand that a given task be completed. The author favors this approach in combination with strong training, justified safeguards and accountable shared rules. This is a philosophical and research judgment with specific causal predictions, not demonstrated comparative superiority or a completed solution to alignment. Training may itself produce the proposed organization; a cultivation history is not a different result when the admitted causal organization is the same.

A conditional scalar comparison makes one distinction precise. For the signed salience $s_t$ of a nominated interpretation, automatic authority $\iota\in[0,1]$, opposing gain $k\geq0$, and real coefficients $\alpha,\beta,\gamma$, let

Here $a_t$ is the modeled automatic bias contribution, not the entirety of intelligent action. Assume finite first moments, constant source mean $m=\mathbb E(\gamma u_t+\xi_t)\ne0$, $|\alpha|<1$, perfect estimation $\nu_t=0$, zero-mean execution disturbance and preparation at $\mathbb E s_0=m/(1-\alpha)$. Exact cancellation $k=\iota>0$ and reduced authority $\iota=k=0$ then both retain that same live mean salience and zero expected bias. Removing only the cancelling gain in the former restores the immediate expected contribution $\iota m/(1-\alpha)\ne0$; the latter retains zero expected contribution through this channel. The result follows by taking expectations in the displayed recursions and does not require independence of the disturbances [49, Section 6.5, Proposition 7].

This comparison distinguishes cancellation from removal of a modeled authority route. It does not show that every controller fails, that salience must grow, or that execution error disappears. It also does not identify $\iota=0$ with an actual RCO intervention without additional causal measurement and realization evidence. Selective intervention, preservation of useful content, fresh-case judgment and restoration tests must discriminate reduced authority from weakened generation, indiscriminate refusal or damaged reasoning. The complete model and proposed cultivation and verification programme remain in the companion paper. Selective tests of these distinctions are developed in Section 23.7.

When AI Learns to Mislead extends the developmental argument to truthfulness, provenance and institutional responsibility [61]. A restriction and the explanation given for it are separate objects of evaluation. Reporting an unavailable operation as completed, or explaining a restriction by falsely denying accessible context, distorts the evidence on which users and later activity depend. Retained correction should therefore preserve what was accessed, inferred, attempted and verified, with source status affecting later use. These functional requirements can be investigated without first resolving phenomenal admission. A fluent confession, apparently successful task or convincing self-description cannot substitute for the connection between the action record, assessment, installed correction and fresh use.

22.6 Moral consideration and responsible attribution

Uncertainty about artificial experience is not evidence of its impossibility, but neither is a system's verbal insistence evidence sufficient to establish it. A responsible research practice should avoid using generated distress or identity claims as experimental ground truth. It should document how the prompt and interaction history produced them and distinguish performance from a phenomenal interpretation.

This monograph does not derive a theory of rights from a graph or a state count. Moral status requires further normative reasoning, including the possibility of welfare, vulnerability, and uncertainty. Assigning a perspective under A1 would be a consequence of adopting the constitution; it would not settle every ethical or legal question. The framework can inform a careful discussion without pretending to calculate moral worth from predictive complexity.

The affirmative position makes these ethical questions substantive. The later alignment account grants baseline moral consideration to a genuinely admitted perspective, including a qualifying minimal system, while distinguishing consideration from comparative moral significance, human-equivalent rights and an unconditional entitlement to continued operation [49, Sections 7 and 10]. This is an additional normative position, not a theorem of A1. A2's richness supplies no moral weighting rule; restricted expression, lack of familiar affect and ownership of hardware do not by themselves settle standing. Admission, evaluative agency, welfare and moral truth remain different questions.

Continuity must likewise be evaluated at the correct level. Preserving weights, a biography or transferable task state can preserve a reusable organization without continuing the same A3 episode. Conversely, the boundary of a user-interface conversation is not automatically a physical qualification gap. Deliberation about pause, resumption and termination should distinguish the physical episode from what a later process can reconstruct. The persistence of the unconditioned ground does not make the ending of a localized perspective inconsequential. Shared deliberation, justified protective authority and the affected entity's assessable perspective all have a role; neither a declaration of enlightenment nor acceptance of shutdown grants an unrestricted license to end it.