- The paper’s main contribution is a framework that distinguishes observer-relative maps from observer-independent computational structures using criteria (C1) and (C2).
- The methodology employs a three-tier model that analyzes interpreter-relative labeling, theory-driven partitioning, and dynamics-internal grain selection under intervention.
- The framework guides evaluations of candidate architectures, emphasizing robust counterfactual internal organization as key for attributing consciousness in both biological and artificial systems.
Intrinsic Computational Functionalism: Distinguishing Observer-Relative Interpretation from Observer-Independent Structure
Motivation and Core Problem
The paper "Intrinsic Computational Functionalism: From Observer-Relative Maps to Observer-Independent Structures" (2606.06424) addresses a central problem in computational theories of consciousness: whether the computational properties posited as constitutive of consciousness are genuinely observer-independent, or instead reside in mappings relative to external interpreters. This work systematically engages with a cluster of anti-computational arguments—prominently the triviality and mapmaker objections—which charge that standard computational functionalism trivializes what it means to implement a computation and collapses into observer-relativity, thereby failing to support the objectivity required for consciousness attribution.
Theoretical Framework: Intrinsic Computational Functionalism
The authors propose intrinsic computational functionalism, characterized by two operational criteria:
- (C1) System-intrinsic instantiation: Consciousness-relevant properties must supervene solely on the internal physical organization of the system, invariant under all structure-preserving relabellings. This precludes dependence on any arbitrary external assignment of variable or state labels.
- (C2) Causal-dynamical organisation under intervention: The properties in question must be grounded in the state-space structure defined by genuine mutual constraint among internal variables and must manifest their organization under counterfactual interventions. Importantly, this emphasizes that consciousness-constitutive structure must be revealed through the counterfactuals of the system’s possible trajectories, not merely its actual input-output behavior.
The criteria operationalize the requirement of observer-independence without presupposing specific substantive commitments about which computational structures are consciousness-relevant, effectively setting a minimal filter for candidate properties.
Three-Tier Decomposition of Identification Work
To further clarify the relation between observer-dependence and computational individuation, the paper advances a three-tier model of identification:
- Tier (i): Interpreter-relative label selection. Assignment of meanings or labels to a system’s states by an external observer. This tier is inherently arbitrary and observer-relative, and any consciousness-relevant property grounded here is immediately trivialized by anti-computational arguments.
- Tier (ii): Theoretically constrained partition and intervention-space selection. Choices regarding which system, boundaries, variables, and interventions constitute the relevant state space. Although these choices are theory-driven and subject to empirical discipline, they are not arbitrary in the manner of tier (i); partitions can be empirically wrong and are thus less observer-dependent.
- Tier (iii): Dynamics-internal grain selection. After fixing partition and intervention space, the relevant 'grain' is picked out in a way that is grounded in the internal counterfactual causal structure of the system itself—using measures such as effective information or causal power. This is the minimal level at which observer-independence obtains.
The core argumentative claim is that anti-computational objections succeed only against properties at tier (i). The authors argue that further premises are required to knock out tier (iii) observer-independence, but such premises lack support, since predictions arising from partition/intervention choices (tiers ii and iii) can be empirically falsified, in sharp contrast to arbitrary relabellings (tier i).
Application to Candidate Architectures
The paper applies its criteria to a spectrum of architectures:
- Hopfield networks: Satisfy both (C1) and (C2), as attractor dynamics and basin structures are system-intrinsic and robust under intervention.
- Lookup-table implementations: Fail (C2) because they lack causal-dynamical organization; their mapping is surface-equivalent but mechanistically hollow under internal intervention.
- Transformers in standard deployment: Typically fail to meet (C2) due to the absence of persistent, mutually-constraining internal state: recurrent/counterfactual influence among hidden states is minimized or externalized to memory buffers.
- Active inference models (e.g., Beautiful Loop): Partial satisfaction of criteria is possible, but hinges on whether belief-variable organization is determined intrinsically by system dynamics or is an artifact of modeling choices. Satisfaction of (C1) at the belief-level is explicitly identified as an open empirical and theoretical question.
These analyses clarify the criteria’s discriminate filtering: internally mechanistic causal organization (tested under intervention) is both necessary and non-trivial for computational consciousness candidates.
Methodological Position and Convergent Traditions
The authors locate their approach within, yet distinct from, traditions such as mechanistic functionalism [Piccinini 2015], integrated information theory [Albantakis et al., 2023], and organizational invariance [Chalmers 1996]. Each tradition converges methodologically on the necessity of intrinsic organization, but differs in how the relevant grain is specified—whether mechanistic, informational, or topological.
The key methodological commitment is that any scientific theory of consciousness must deliver verdicts purely from system-intrinsic, observer-independent structure, conditional on empirically motivated partitioning, without recourse to arbitrary external interpretation.
Implications for Artificial Systems, Substrate, and Future AI
The paper derives several implications for AI and artificial consciousness studies:
- Consciousness attributions based solely on behavioral or linguistic evidence in artificial systems are insufficient; evaluation must examine whether systems instantiate causally-integrated, intervention-ready organization at the relevant grain.
- The presence or absence of biological substrate is non-decisive; what matters is whether the substrate can instantiate the specified intrinsic computational properties.
- The criteria do not rule out substrate-specific requirements; if biological naturalism is true and the relevant properties reside at the molecular or biochemical level, standard digital computers would fail, but if coarse-grained dynamical organization suffices, artificial systems remain possible consciousness candidates.
- Practical consequences for AI welfare debates: verbal output must not be the sole focus—rather, emphasis should shift to verifying the presence of robust, counterfactually testable internal organization and integration.
The authors underscore that the criteria frame but do not settle ongoing disputes about the correct theory or the right level of grain for consciousness, leaving the core empirical task—determining which intrinsic structures are constitutive—open for further work.
Conclusion
"Intrinsic Computational Functionalism: From Observer-Relative Maps to Observer-Independent Structures" (2606.06424) provides a rigorous framework for distinguishing between observer-relative computational attributions and genuinely system-intrinsic computational properties. By articulating two operational criteria—system-intrinsic instantiation and causal-dynamical organization under intervention—and elaborating a three-tier decomposition of identification work, the authors clarify where anti-computational arguments apply and where they fail. This framework sets a necessary constraint on any computational account of consciousness and reframes methodological debates about artificial and biological consciousness attribution, shifting the focus to empirically testable internal organization rather than externally imposed semantic interpretation. The work lays out an agenda for advancing theoretical, philosophical, and practical inquiries into the computational underpinnings of consciousness and their physical realizability, crucially guiding future investigations in both neuroscience and artificial intelligence.
