Malinkovic and Aru’s Biological Computationalism

Neuroscientists Borjan Malinkovic and Jaan Aru propose a theory of consciousness they call “biological computationalism,” a framework where the specific physical and energetic properties of biological substrates are not merely implementation details, but constitutive of the computations capable of supporting consciousness. They argue that “biological systems support conscious processing because they (i) instantiate scale-inseparable, substrate-dependent multiscale processing as a metabolic optimisation strategy, and (ii) alongside discrete computations, they perform continuous-valued computations due to the very nature of the fluidic substrate from which they are composed.” They claim that “these features—scale inseparability and hybrid computations—are not peripheral, but essential to the brain’s mode of computation” (Malinkovic and Aru, 2025).

Thus, Malinkovic and Aru challenge the general orthodoxy of “computational functionalism,” the oft-default philosophical stance that mental states supervene solely on the right kinds of patterns of information processing, irrespective of the physical medium in which they are instantiated. They argue that the distinction between biological and digital systems is foundational, rooted in the architectural separability of the latter. Contemporary deep learning systems—Large Language Models (LLMs)—have strict separations of scales and components. In these systems, memory, processing (the arithmetic-logic unit), and control are physically isolated, communicating via buses. The software algorithms are distinct from the hardware implementation. In contrast, biological computation is characterized by substrate dependence, in which the physical dynamics of the system—membrane potentials, ion flows, and chemical diffusion—are the computational substrate. In the brain, there is no "hardware abstraction layer"; the algorithm and its realization are co-instantiated and inseparable. As the authors pithily state: "The substrate is not merely a carrier of the algorithm—the substrate is the algorithm” (Malinkovic and Aru, 2025).

Metabolic Constraints

A foundational driver of biological computationalism is that the brain’s organizational structure is a direct evolutionary response to intense and unrelenting metabolic constraints. Unlike artificial neural networks, which can scale performance simply by consuming more energy—adding parameters and compute cycles at will—the brain operates under a strict caloric budget, consuming roughly 20 watts. This energy limitation, or scarcity, constrains the nervous system to adopt specific processing strategies that optimize computational capacity per unit of energy. This required optimization necessitates that information be aggregated from microscopic discrete events into macroscopic continuous dynamics to reduce costs. For example, non-spiking neurons use graded potentials to relay signals, capable of carrying significantly more information per second than metabolic-heavy spiking elements. The authors argue that these energy constraints actively shape the system's topology: "Metabolic constraints are therefore constitutive forces, shaping how neural systems process, transmit, and integrate information across scales" (Malinkovic and Aru, 2025).

In the authors’ words, “The brain’s response to metabolic constraints—that of coarse-graining information across scales and retaining multiscale interdependencies—creates a particular form of organisation fundamentally different from engineered hierarchies.” The result is a neuronal system where computations at one scale cannot be cleanly separated from those at another. The authors call this multiscale interdependency a "heterarchy," to distinguish it from the rigid hierarchies of engineered systems. “This reveals a formal notion of scale-integration that assumes heterarchy over hierarchy: proving statements at one scale requires simultaneous access to information from other scales, much like Tarski hierarchies, but with continuous co-determination rather than unidirectional queries.”

Comparing Digital and Biological Scales and Structures

To be specific, in digital architectures, levels of abstraction are unidirectional; a higher-level coding language does not physically alter the transistor structure and logic beneath it in real-time. Biological systems, however, exhibit scale-inseparability, meaning there is no single privileged scale of dynamical processes that underpins consciousness. This undergirds the authors’ sophisticated analogy to Tarski hierarchies in formal logic, where truth at one level requires a meta-level for definition. Similarly, biological systems utilize a multiscale structure where lower scales endogenously generate higher scales, while higher-scale dynamics simultaneously constrain the activity of the lower scales.

The authors cite evidence for “diffuse, continuous processes at the subcellular level exerting profound effects on network-level transitions,” which extend our understanding of graded potentials. These processes can be considered emergent temporal coarse-grainings, or slow features, that evolve over longer intervals than fast electrophysiological activity. A specific example is Protein Kinase A (PKA), which “compares internal activity with that of postsynaptic neurons, functioning as a decision process or a continuous evidence accumulation mechanism determining when the network’s dynamic goal has been reached,” which then triggers discrete network-level transitions (Malinkovic and Aru, 2025).

This phenomenon is described as "dynamico-structural co-determination." It’s a mutual dependency implying that consciousness may require a specific type of emergent closure across scales that digital simulations—which model these scales as distinct mathematical objects—cannot natively replicate. The authors emphasize this crucial difference: "Computations [in the brain] at any one scale are not separable from each other... This scale inseparability suggests that there is no single privileged scale of dynamical processes that underpin consciousness."

The authors stress that this sharp contrast between brain-based and chip-based computations highlights “an important conceptual caution: importing metaphors from engineered systems (which assume clean hierarchies) can obscure the functional organisation of biological brains. “We require frameworks accounting for multiscale integration, dynamico-structural co-determination, and the fluid, heterarchical nature of evolved neural processes” (Malinkovic and Aru, 2025).

Hybrid Nature of Neural Dynamics

The authors further differentiate biological processing through its hybrid nature, which integrates discrete signaling with continuous physical dynamics. While the long-time, standard neuron model—"integrate, trip threshold, fire"—describes a digital-like operation of summing inputs to a threshold, biological reality is far more complex. Neural tissue utilizes continuous-valued processes, such as graded membrane potentials and extracellular electric fields, which work non-digitally in an analogue regime of "continuous arithmetic" that is formally distinct from the discrete symbol manipulation of Turing machines.

Dendritic computation exemplifies this hybridity. Dendrites utilize active voltage-gated channels to perform non-linear operations locally within the branchings, temporally aggregating (coarse-graining) inputs before they ever reach the soma to trigger (or not) the neuronal spike. Furthermore, neurons communicate through ephaptic coupling, where local electric fields modify the excitability of neighboring cells without synaptic contact, creating a "field-like" medium of information transfer that is spatially continuous. These oscillatory fields are not epiphenomenal waste products, rather information-bearing factors that influence, guide, and synchronize discrete neuronal firing events . The authors describe this interplay: "Spikes (discrete events) ride atop graded membrane potentials, ion flows, and field potentials (continuous processes), which themselves ride atop layer-integrated cortical anatomy" (Malinkovic and Aru, 2025).

Implications for Artificial Consciousness

A clear consequence of biological computationalism as a theory of consciousness would dampen the “prevailing optimism” that artificial systems, particularly LLMs, have the capabilities to one day be conscious. To reiterate, Malinkovic and Aru contrast computational functionalism (i.e., consciousness is solely the right pattern of information processing, independent of the physical substrate) with biological naturalism (i.e., conscious experience is dependent on the concrete physical processes of living systems). Despite the centrality of these positions to the artificial consciousness debate, the authors make the critical point that “there is currently no coherent framework that explains how biological computation differs from digital computation, and why this difference might matter for consciousness.” They then argue that “the absence of consciousness in artificial systems is not merely due to missing functional organisation but reflects a deeper divide between digital and biological modes of computation and the dynamico-structural dependencies of living organisms” (Malinkovic and Aru, 2025).

The authors contend that because current AI systems, including LLMs, are built on von Neumann architectures that decouple information from its physical substrate, they fundamentally lack the causal structures necessary for consciousness. The missing ingredient is not merely algorithmic complexity but the presence of living-systems-based computation—operations grounded in finite, substrate-specific resources where the system’s existence is tied to its thermodynamics.

To achieve true artificial consciousness, the authors propose moving beyond silicon-based digital logic toward biomimetic substrates that naturally support hybrid, multiscale processing. Such devices would require intrinsic "structured stochasticity" and memory that depends on physical history, akin to biological synapses. "Systems capable of supporting consciousness must integrate energy, structure, and dynamics, not merely manipulate symbols." Indeed, machines would have to embody the messy, energy-constrained, and inseparable nature of life itself (Malinkovic and Aru, 2025).