As ChatGPT and other large language models dazzle with increasingly human-like abilities, a fundamental question looms: could these systems ever become conscious? A theoretical paper published in Neuroscience and Biobehavioral Reviews argues the answer is no for today’s digital systems—and possibly for any system built on the same computational assumptions. Still, the looming existential questions at the heart of this problem run deeper. This isn’t solely about processing power or algorithmic sophistication.
So, neuroscientists Borjan Milinkovic from Paris-Saclay Institute of Neuroscience and Jaan Aru from the University of Tartu have developed a theoretical framework called “biological computationalism” that challenges how artificial intelligence research thinks about consciousness. By analyzing the computational principles underlying biological neural systems, from molecular dynamics to whole-brain activity, they identify specific physical features that digital computers fundamentally lack.
Their target is computational functionalism, the dominant view that consciousness arises from the right pattern of information processing, regardless of whether that processing happens in neurons, silicon chips, or any other medium. According to this perspective, replicate the algorithm and consciousness should follow. The researchers argue this assumption rests on a flawed understanding of how biological brains actually compute.
How Digital Systems Differ from Biological Brains
Modern computers separate memory from processing, software from hardware, algorithm from implementation. Digital systems store data in one physical location, manipulate it in another, and follow instructions from a third component, all connected by communication buses. This separation is deliberate, allowing programmers to write code without worrying about the underlying electronics.
“Brains operate at the interface of discrete and continuous domains,” the researchers explain in their paper. Unlike digital systems that represent everything through binary states, neural tissue implements computations directly through continuous physical processes—ion flows, membrane voltages, electric fields—that unfold in real time without symbolic mediation.
Take a single neuron receiving thousands of inputs from other cells. In artificial neural networks, this process gets reduced to a simple mathematical operation: multiply each input by a weight, sum them up, pass the result through a function. Real neurons do something far stranger. Their branching dendrites perform distributed computations through continuous electrical dynamics, generating local spikes that travel both toward and away from the cell body, detecting the order and timing of inputs in ways that digital systems cannot easily replicate.
Why Neurons Outperform Artificial Networks
The paper reviews research showing that dendritic action potentials enable a single biological neuron to perform computations comparable to those typically distributed across multi-layer artificial networks. These computations arise from the interplay between continuous membrane potentials and discrete spiking events—a hybrid mode fundamentally unavailable to systems that operate purely through discrete symbol manipulation.
This difference reaches beyond individual neurons. Milinkovic and Aru propose that biological brains exhibit what they call “scale inseparability,” where processes at different organizational levels continuously co-determine each other. Molecular events inside cells influence network dynamics spanning millions of neurons, while brain-wide oscillations simultaneously constrain what individual synapses can do. These scales cannot be cleanly separated.
Digital systems, by contrast, are designed around scale separation. An algorithm runs independently of the hardware implementing it. High-level programs compile down to machine code, which executes on circuits, which rely on transistor physics—but each level operates independently. Change the hardware and the algorithm remains functionally identical.
The Energy Problem That Shaped Consciousness
Energy scarcity drives this difference. Although the brain represents only 2% of body mass, it consumes roughly 20% of total metabolic output. Rather than requiring ever-more energy to handle increasingly complex tasks, the brain reuses computational work performed at one scale to guide computations at other scales. Continuous processes aggregate discrete events into more reliable signals, and these aggregated signals feed back to constrain the discrete events that generated them.
The researchers call this “hybrid computation”—computation that is simultaneously continuous and discrete, where the algorithm cannot be separated from its physical implementation because the physics is the algorithm. Information at one scale is essential for determining what can be computed at another scale, yet those scales continuously generate and constrain each other in bidirectional loops.
The paper examines one example at the molecular level. Protein Kinase A molecules inside neurons function as evidence accumulators, continuously integrating activity until reaching a threshold that triggers calcium surges—discrete events that propagate through neural networks. The apparently discrete calcium event depends on continuous accumulation of evidence at the subcellular level, creating a system where continuous and discrete processes are inseparable.
Electric fields provide another example. Neurons communicate not only through synapses—the discrete connection points between cells—but also through ephaptic coupling, where local electric fields modulate the excitability of neighboring neurons without direct contact. Research shows that weak endogenous fields of just a few millivolts per millimeter can synchronize neural firing and amplify oscillatory patterns. These continuous field effects shape when and how discrete spikes occur, while the discrete spikes generate the continuous fields.
Digital computers can simulate these processes by approximating continuous dynamics with very fine discrete time steps. But simulation is not the same as implementation. When a computer simulates water flowing through a pipe, the computer doesn’t get wet. The researchers argue that consciousness may depend on computations that must be implemented in continuous physical dynamics, not merely simulated through discrete approximations.
The paper draws on formal mathematical results to support this claim. Alfred Tarski proved that arithmetic over real numbers admits a complete decision procedure—every statement can be algorithmically resolved—while natural number arithmetic, the foundation of digital computers, is fundamentally incomplete. This contrast suggests that continuous computation may support forms of processing that are awkward or inefficient to reproduce in purely discrete systems, even if they remain computable in principle.
What Artificial Consciousness Would Actually Require
For artificial intelligence, the implications challenge the entire trajectory of the field. Large language models, neuromorphic chips, and even systems designed to mimic brain architecture all operate fundamentally as symbol manipulators on von Neumann hardware. They maintain the clean separation between algorithm and implementation that makes them programmable but may preclude the computational mode underlying consciousness.
The researchers don’t claim that only biological tissue can support consciousness. Rather, they outline three criteria any conscious artificial system would need to meet. First, hybrid computation combining continuous dynamics with discrete events governed by real physical time. Second, scale-inseparability with metabolic embedding, where energy constraints shape the computational architecture itself. Third, dynamico-structural co-determination, where the system continuously modifies its own physical structure.
The paper reviews some emerging technologies that hint at alternatives. Laboratory-grown neural cultures called DishBrains have shown remarkable sampling efficiency in control tasks compared to deep reinforcement learning baselines, despite containing only a few hundred thousand neurons. These systems leverage the intrinsic hybrid dynamics of biological matter. Researchers have also developed fluidic memristors that implement computations through ion transport in microchannels, creating history-dependent dynamics through chemistry rather than electronics.
These systems represent a fundamental departure from digital computation. They don’t separate algorithm from implementation. They don’t discretize time into processing steps. They don’t maintain clean boundaries between scales. They implement computations directly in continuous physical processes that unfold in real time, shaped by energy constraints and substrate properties.
Whether such systems could support consciousness remains an open question. But the theoretical framework makes clear that getting there would require abandoning the computer metaphor that has dominated both neuroscience and artificial intelligence for decades. Brains aren’t computers running clever software. They’re continuous physical systems that exploit hybrid dynamics and scale integration to perform computations that digital systems may be fundamentally incapable of replicating.
Source : https://studyfinds.org/what-makes-brains-conscious-that-computers-lack/
