Intrinsic Computational Functionalism

• Intrinsic Computational Functionalism is a view that defines consciousness through observer-independent, physically realized computational structures.
• It emphasizes invariance under relabeling and counterfactual dynamics to filter out superficial behavioral equivalence.
• The framework employs canonical functionalism and mechanism-enriched realization to formalize consciousness-relevant computational properties.

Searching arXiv for papers on intrinsic computational functionalism and closely related frameworks. {"query":"all: \"intrinsic computational functionalism\" OR \"canonical functionalism\" consciousness", "max_results": 10} I found several recent arXiv entries relevant to intrinsic computational functionalism, including work on observer-independent computational structure, canonical functionalism, simulated consciousness, and critiques from IIT and biological naturalism. I'll use those alongside the provided paper set. Intrinsic Computational Functionalism (ICF) is the view that, if consciousness is computationally constituted, it depends on physically realised computational structures that a system has in virtue of itself rather than on labels imposed by an external interpreter. In this literature, the central contrast is between observer-relative computational descriptions and observer-independent organization grounded in the system’s own causal-dynamical structure. ICF is therefore an anti-interpretivist refinement of computational functionalism: it does not treat arbitrary semantic mappings, superficial behavioral equivalence, or mere actual input-output matching as sufficient. Instead, it asks which computational properties could remain invariant under relabelling, intervention, and changes of realization while still being genuinely instantiated by the system itself (Ma et al., 4 Jun 2026).

1. Core conception and philosophical target

The immediate target of ICF is the family of objections according to which computation, when used in consciousness theory, is too observer-relative to do explanatory work. On this diagnosis, the same physical process can be assigned incompatible computational interpretations by changing the labeling scheme, so no externally imposed mapping can ground consciousness. ICF accepts that criticism against interpretation-based computation, but denies that it generalizes to all computational organization. Its claim is narrower and more demanding: only computational structure that is physically realized by the system’s own organization can qualify as consciousness-relevant (Ma et al., 4 Jun 2026).

Recent formulations operationalize this position with two necessary criteria. (C1) System-intrinsic instantiation requires that the relevant property be specifiable without an observer’s labelling and invariant under structure-preserving relabellings of the system’s variables. (C2) Causal-dynamical organisation under intervention requires that the property be grounded in a state-space structure whose variables mutually constrain one another, and whose organization is exhibited in counterfactual response under intervention. Together, these criteria exclude arbitrary semantic projection, recorded replay, and lookup-table style mimicry, while leaving open which intrinsic structures, if any, are sufficient for consciousness (Ma et al., 4 Jun 2026).

This anti-interpretivist orientation has antecedents in earlier invariance-based defenses of functionalism. A Level-1 functionalist framework was defined by functional states individuated purely structurally, with prediction depending only on functional state; within that framework, no Type-2 variation exists, and hence no substitution exists, for a broad class of Level-1 functionalist theories. The result supports the claim that consciousness-relevant prediction should be invariant across differences in encoding or representation when functional organization is preserved (Ganesh, 2020).

2. Canonical functional structure

A mathematically precise refinement of this program is canonical functionalism, which replaces vague appeals to “functional organization” with a canonical quotient structure. The system is modeled as a deterministic interactive Moore machine

$S=(X,x_0,\delta,o),$

with internal state space $X$, initial state $x_0$, transition function $\delta:X\times I\to X$, and output function $o:X\to O$. For each internal state $x$, its complete future behavior is

$b_x^S:I^*\to O, \qquad b_x^S(w)=o(\delta^*(x,w)).$

Two states are equivalent iff they have identical future behavior under all possible continuations: $$x\sim_S y \quad \text{iff} \quad \forall w\in I^*,\; b_x^S(w)=b_y^S(w).$$ If $R_S$ is the reachable state set, the canonical functional structure is the quotient

$\mathrm{Can}(S)=R_S/{\sim_S}.$

The induced transition and output maps are well-defined on equivalence classes, and if two systems have the same complete behavior, $X$0, then their canonical structures are isomorphic: $X$1 The construction is explicitly described as a Myhill–Nerode / minimal automaton style construction, adapted to interactive systems and then philosophically repurposed for consciousness (Kanai et al., 9 May 2026).

The importance of this framework for ICF is methodological. It avoids observer-relative semantic maps because states are not individuated by imposed meanings such as “this voltage represents 1” or “this state means red.” They are individuated by complete counterfactual role: how the system would evolve and respond from that state under all possible future interactions. Once the interface is fixed, the canonical structure is determined by the system’s own counterfactual transition organization. This does not identify which systems are conscious, nor does it show that functional organization is sufficient for consciousness. Rather, it identifies the canonical object over which a functionalist theory should be formulated: consciousness-relevant properties should factor through the quotient, so that if $X$2, then the consciousness-relevant property should be the same for $X$3 and $X$4 (Kanai et al., 9 May 2026).

This shift reframes classical objections. Lookup tables, simulations, and unfolded realizations do not by themselves refute functionalism; they force the theorist to specify whether the relevant canonical structure is preserved. At the purely boundary-functional level, if the complete counterfactual organization is preserved, the systems are canonically equivalent. A plausible implication is that canonical functionalism marks a behavioral boundary case of ICF rather than its final form.

3. Mechanism-enriched intrinsic structure and realization

A later extension argues that boundary-level canonical structure is too weak for consciousness because it can collapse systems that share the same external input-output behavior but differ in internal mechanism. On that view, a consciousness-relevant canonical representation must include internal mechanisms, interventions, and joint readouts belonging to the relevant intrinsic organization. The resulting formalism models a physical system as

$X$5

where $X$6 is the physical state space, $X$7 the admissible input/context continuations, $X$8 the admissible interventions, $X$9 a family of admissible readout sets, $x_0$0 the corresponding joint readouts, and $x_0$1 the time evolution under intervention $x_0$2 and context $x_0$3. For $x_0$4, the relevant role is no longer just complete boundary behavior; it is the full profile of readout trajectories across admissible interventions, contexts, and internal or joint readouts (Kanai et al., 13 Jun 2026).

Within this framework, a partition of state space counts as intrinsic only if it satisfies five conditions: physical support, intervention tracking, relabelling invariance, causal-dynamical individuation, and dynamical closure or lumpability. Given such a partition $x_0$5, the intrinsic causal-computational structure is represented as

$x_0$6

with $x_0$7 the induced transition structure and $x_0$8 the induced counterfactual response profile. If the partition is the finest one induced by mechanism-enriched equivalence, this yields a mechanism-enriched canonical structure, $x_0$9. The earlier canonical functional structure is then recovered as a special case in which the only readout is external output and the only interventions are ordinary input histories (Kanai et al., 13 Jun 2026).

The associated realization relation is Intrinsic Causal-Computational Realization (ICCR), written

$\delta:X\times I\to X$0

ICCR requires physical realization, transition preservation, counterfactual preservation, and boundary adequacy. The central theorem is conditional: if a conscious property $\delta:X\times I\to X$1 is grounded in intrinsic causal-computational organization, and if a candidate system satisfies ICCR with respect to a target, then

$\delta:X\times I\to X$2

This makes the simulated-consciousness dispute structurally explicit. A system’s being “a simulation” is not, by itself, disqualifying; to deny consciousness, one must identify a consciousness-relevant intrinsic causal-computational structure that the candidate fails to realize (Kanai et al., 13 Jun 2026).

The motivating examples are lookup tables and unfolded systems. At the boundary level, a recurrent system and an unfolded feedforward system can be equivalent. Once internal interventions and joint readouts are included, they can diverge because there is no structure-preserving bijection for the relevant persistent internal variables or intervention-sensitive mechanism. The criticism is therefore not that simulation is impossible in principle, but that boundary equivalence is too coarse whenever consciousness depends on internal organization.

4. Intrinsic computation, programmability, and organised matter

A distinct but related line of argument links computational functionalism to mortal computation rather than ordinary Turing computation. Starting from a formal restatement of computational functionalism—there exists at least one computation $\delta:X\times I\to X$3 such that every system capable of a relevant experience realizes $\delta:X\times I\to X$4—the argument adds Assumption 1, namely that there exists at least one conscious system that is not capable of running ordinary programs. Programs are defined as a paradigmatic subclass of immortal computation, and every Turing computation is taken to be implementable by some program. Under these assumptions, the consciousness-relevant computation cannot be a program, and therefore cannot be a Turing computation. Since mortal computation is defined as the complement of immortal computation, consciousness, if computational, must be a mortal computation: a computation whose realizability depends on the kind of hardware that performs it and whose identity may depend on intrinsic physical properties such as variations in connectivity or nonlinearities (Kleiner, 2024).

This argument does not prove that artificial systems cannot be conscious. The conclusion is explicitly weaker and more structural: consciousness, under computational functionalism plus Assumption 1, cannot be identified with a portable program or with an abstract Turing computation. The paper also notes that “mortal computation” admits both an epistemic reading, on which it is simply a Turing computation not fully known by an outside programmer, and an ontic reading, on which it exceeds Turing-style constraints. The formal result requires only the weaker point that consciousness falls on the mortal side of the immortal/mortal divide (Kleiner, 2024).

Another related reformulation defines computation as an organizational property of matter rather than as execution of an abstract algorithm over symbolic representations. On that account, organization is the persistence of relational constraints delimiting admissible state transitions; information is relational invariance, meaning a difference that alters which future transitions remain admissible; and computation is the ongoing enactment of such organization. Memory, processing, and execution are not separate stages but inseparable aspects of material dynamics. The framework proposes experimentally oriented criteria for identifying computation: persistence of relational invariants, recovery after perturbation, structural failure under perturbation, and phase boundaries in constraint–energy space (Witte, 7 Jan 2026).

These developments do not collapse into a single doctrine. The mortal-computation argument emphasizes hardware dependence and the limits of programmability; the organizational account emphasizes persistent relational constraints and perturbation-based criteria. What they share with ICF is rejection of the idea that consciousness-relevant computation is exhausted by externally specified software, abstract syntax, or input-output matching alone.

5. Critiques, alternatives, and rival constraints

The most direct rival to ICF in this literature is Integrated Information Theory (IIT). An IIT-based argument claims that two systems can be functionally equivalent without being phenomenally equivalent because consciousness depends on intrinsic cause-effect structure rather than on computation alone. In the central example, a four-unit Boolean target system, PQRS, is a complex with $\delta:X\times I\to X$5 intrinsic bits in state $\delta:X\times I\to X$6, with 13 distinctions, 8184 relations, and $\delta:X\times I\to X$7 ibits. A 117-unit stored-program computer can simulate the same state transitions, yet at the micro grain the computer fragments into 24 individual complexes, the whole computer has $\delta:X\times I\to X$8, and no valid macroing that preserves the simulation function recreates the target’s cause-effect structure. The conclusion is that functional equivalence does not imply phenomenal equivalence (Findlay et al., 2024).

This IIT-based challenge does not simply deny that intrinsic organization matters; it proposes a more specific intrinsic organization than ICF itself presupposes. Recent ICF work treats IIT as both ally and opponent: ally, because both reject observer-relative computation; opponent, because IIT typically fixes consciousness to a particular intrinsic cause-effect structure and often uses exclusion or maximal integration principles that ICF does not assume (Kanai et al., 13 Jun 2026).

The anti-functionalist use of IIT has itself been criticized. A gradual neuronal replacement argument against the claim that human-level artificial intelligence implemented on conventional computing hardware is necessarily not conscious pressures the anti-functionalist to explain why consciousness should disappear while behavior, cognition, and self-ascription remain unchanged. The critique does not reject IIT outright, but argues that strong metaphysical claims about substrate-specific impossibility outrun what the theory presently justifies (Shanahan, 2015).

Biological naturalism introduces a different constraint. A recent distinction separates Type-A-BN, on which biology intrinsically matters for consciousness without affording unique information processing capabilities, from Type-B-BN, on which biology matters because it affords unique information processing capabilities. Type-A-BN is argued to be empirically untestable because it dissociates consciousness from behavior and other observable consequences. Type-B-BN is testable, and not incompatible with computational functionalism, because it must specify both what biological feature is unique to biology and why that feature matters for consciousness. On this view, biology can guide inquiry but cannot replace the task of relating consciousness to information processing (Klatzmann et al., 1 Jun 2026).

A broader taxonomy places these disputes across Marr-like levels. Some objections target input-output computability, some target algorithmic organization, and some target physical implementation. The most direct challenges to ICF are level-3 objections invoking causal structure, analog processing, electromagnetic fields, slicing problems, or substrate-specific properties beyond medium-independent computation (Campero et al., 20 Nov 2025).

6. Methodological role and unresolved questions

ICF is not, in these papers, a completed theory of consciousness. It is a constraint on what a computational theory would have to look like if it is to remain observer-independent. Canonical functionalism explicitly states that it does not identify which systems are conscious, nor show that functional organization is sufficient for consciousness. Instead, it identifies the formal object over which such a theory should be stated, and requires consciousness-relevant properties to be invariants over canonical structures rather than artifacts of arbitrary semantic interpretation (Kanai et al., 9 May 2026).

This methodological role becomes more explicit in the three-tier decomposition of identification work. Tier (i) is interpreter-relative label selection; tier (ii) is theoretically constrained partition selection; tier (iii) is dynamics-internal grain selection. Anti-computational arguments succeed against tier-(i) accounts, where computational content is imposed by external labeling. ICF holds that any candidate consciousness-relevant computational property must be identified, if at all, through tier-(iii) dynamics-internal grain selection, conditional on empirically disciplined tier-(ii) choices (Ma et al., 4 Jun 2026).

The main open problem is therefore not whether observer-relative computation is inadequate; that point is largely granted. The unresolved issue is which intrinsic structures, measures, or invariants should be taken as consciousness-relevant. Canonical functionalism lists recurrent organization, global availability, temporal continuity, self-modeling, causal integration, representational geometry, closure or autonomy, and memory and counterfactual responsiveness as the kinds of properties a functionalist theory might specify over $\delta:X\times I\to X$9, while declining to treat any of them as settled (Kanai et al., 9 May 2026).

A second unresolved issue concerns substrate dependence. The mortal-computation argument suggests that consciousness may require hardware-specific realization; IIT argues that intrinsic cause-effect power rather than abstract function is decisive; biologically informed approaches argue that biology matters only if it affords unique information processing capacities. These positions do not converge on a single criterion. What they share is a rejection of naive implementation-independence and a demand that consciousness-relevant organization be identified in the system itself rather than in an observer’s description.

In that sense, Intrinsic Computational Functionalism names less a final doctrine than a restricted thesis: if consciousness is computationally constituted, the relevant computation must be observer-independent, physically realized, intervention-sensitive, and anchored in the system’s own causal-dynamical organization. Whether that requirement is ultimately captured by canonical functional structure, mechanism-enriched realization, mortal computation, intrinsic cause-effect structure, biologically distinctive processing, or some further formalism remains an open question in the current literature (Ma et al., 4 Jun 2026).