2026-09-19
Milinkovic and Aru: consciousness needs hybrid, scale-inseparable, self-rewiring computation; von Neumann LLMs fail all three, fluidic or iontronic hardware is closer.
Once large language models started talking fluently, machine consciousness left the seminar room and entered engineering arguments. Computational functionalism says the right pattern of information processing is enough, whatever the hardware. Biological naturalism says conscious experience is tied to the concrete physics of living tissue, so a digital machine is the wrong kind of object.
Both sides have a piece of the puzzle. What has been missing is a clear account of how biological computation differs from digital computation, and why that difference would matter for consciousness. Borjan Milinkovic at the Paris-Saclay Institute of Neuroscience (NeuroPSI, CNRS) and Jaan Aru at the University of Tartu fill that gap with a 2026 review in Neuroscience & Biobehavioral Reviews, under the label biological computationalism.
The claim is structural, not cosmetic. Digital systems and evolved nervous systems compute in different modes. Biological systems support conscious processing, the paper argues, because they do two things at once. They run scale-inseparable, substrate-dependent multiscale processing as a metabolic optimisation strategy. They also run continuous-valued computation, forced on them by a fluidic substrate, alongside discrete events. Those two features are treated as essential to how brains compute, not as implementation detail.
For any artificial system that wants to approximate consciousness-relevant computation, the paper lists three criteria (Figure 3):
The design choice is to stop treating "more brain-like algorithms" as the next layer of the same stack. A von Neumann machine isolates memory, arithmetic, and control; the instruction set then insulates the algorithm from the transistors. That closure is what classical computability wants, and it is where functionalism gets substrate independence. Brains are not closed that way.
Ion-channel kinetics get tuned to ATP cost. When higher firing rates are needed, neurons more often raise potassium conductance than pack in extra sodium channels, because the cost per spike is lower (Hasenstaub et al., 2010). Structure tracks energy demand; dynamics then rewrite structure. That loop is what the paper names dynamico-structural co-determination.
Continuous and discrete are not two add-ons. Some neurons never spike; they coarse-grain incoming discrete synaptic events with graded potentials. Electric fields and ephaptic coupling modulate excitability without a chemical synapse. Large-scale ionic gradients can press on local populations from the macroscopic end. Dendrites are not passive cables: NMDA receptors and local dendritic spikes make the computation nonlinear and history-dependent before the soma ever sees it. Point neurons that collapse weighted inputs into a sum miss that organisational principle.
On scale, the argument is metabolic. Energy is too tight for clean layers. Processes have to be reused across scales, so a hierarchy compresses into a heterarchy: no privileged scale, no algorithmically closed level. Higher scales can adjudicate computability limits that show up lower down; lower scales generate the higher ones. The coupling is bidirectional, unlike a Tarski meta-level that only looks one way. Unity and differentiation of conscious experience are said to need exactly this kind of cross-scale tethering.
This is a position review. There is no new experiment and no new benchmark. What can be checked is a set of criteria plus a qualitative hardware comparison.
| System | Verdict in the paper |
| Current LLMs on von Neumann machines | Externally powered, metabolism uncoupled from computation, scales separable; fail the three criteria |
| Digital neuromorphic platforms such as SpiNNaker | Discretised simulation, not continuous dynamics in physical time |
| BrainScaleS-2 | Mixed-signal analogue circuits; membrane and synaptic dynamics unfold through capacitors and conductances; a step beyond pure digital platforms, still short of the three criteria |
| Fluidic, iontronic, or synthesised biological tissue | Named as closer to "substrate physics is the computation" |
Figure 3 is a conceptual design path: biologically informed substrates at the base, dynamics unfolding in real physical time under energetic constraint, then scale integration and hybrid event-field interaction. The closing line is blunt. Silicon digital computers, however sophisticated, may be insufficient. The ontology of the computation is wrong. Scaling the model does not fix that.
The three conditions are necessary, not sufficient. Meeting them would not prove consciousness. Artificial consciousness is not ruled out. Scaling the present digital stack is.
For people who score models on consciousness indicators, this paper changes the exam. Checklists that extract properties from neuroscience theories under computational functionalism can miss the substrate layer entirely. Whether an LLM looks conscious can still be measured. Whether it is conscious in the biological sense, the paper says, first requires a different machine.
For hardware, the scoring rule shifts from "does the spiking net resemble cortex" to "can the computation be peeled off the physics." Mixed-signal neuromorphic is a step, not a destination. Fluidic and iontronic devices sit closer to the target than another GPU cluster. That is a hardware roadmap, not a new theory that immediately falsifies functionalism.
The usable artifact is the three-item list. Run it against a system. If it fails, drop the word consciousness.
The authors write the three conditions as necessary, insufficient, and incomplete. The claim that unity, differentiation, temporal coherence, and context-sensitivity "appear to require" this computational regime is an inference, not a lesion study. There is no experiment here that knocks out cross-scale coupling or electric fields and watches consciousness fail.
The BrainScaleS-2 verdict is coarse. Mixed-signal analogue does put dynamics in physical time, but "does not yet meet the tripartite criteria" is not operationalised. A reader cannot take the same ruler to Loihi, TrueNorth, or the next platform.
Metabolic optimisation is offered as the reason scales refuse to separate. Ion-channel evidence supports energy shaping structure. The jump from that fact to consciousness still has several untested steps. Digital systems can wear a power budget as an external constraint; the paper would call that an add-on, not co-determination of computation and structure. The distinction is conceptual, not measured.
The publisher site blocked this host, so the write-up was checked against CC BY excerpts of the published article indexed from the journal page. Individual figure panels and review tables were not compared cell by cell. That gap does not change the numbers, because the paper reports almost none.