An MIT theory of analog computation challenges the computer metaphor—and raises difficult questions about intelligence, consciousness, and AI.
You can map every connection in a machine and still misunderstand what the machine is doing.
That would be an inconvenience if the machine were a server. It becomes considerably more interesting when the machine is the human brain.
For decades, neuroscience has treated the brain the way engineers treat an unfamiliar system: trace the wiring, record the signals, build models that predict outputs from inputs. It has worked remarkably well. It also carries a quiet assumption—that the electrical rhythms sweeping across the cortex are mostly traffic between the parts doing the real work.
In August 2026, three neuroscientists at MIT's Picower Institute—Earl K. Miller, Scott L. Brincat, and Jefferson E. Roy—published a review in The Journal of Neuroscience that questions that assumption. Brain waves, they propose, may not merely carry information between computing elements but take part in the computation—and the same dynamics may help explain unified conscious experience.
It is a serious proposal. It is not a proven one. Both facts matter for what follows.
The metaphor that got comfortable
The computer metaphor earned its place. It gave neuroscience a vocabulary—encoding, memory, computation—and gave artificial intelligence one of its founding inspirations.
But every metaphor is an abstraction, and abstractions decide in advance what counts as detail. A digital computer represents information in discrete states, engineered so that the continuous physics underneath stops mattering: small voltage differences within a logic level all read as the same bit. That indifference is what makes digital systems reliable and portable across hardware.
Brains were not engineered for that indifference. Neurons fire discrete spikes, which look pleasingly digital, but they sit in a conductive medium of continuously varying potentials, chemical signals, and rhythms across many scales. "Digital" and "analog" are not two kinds of brain; they are two descriptions, and one system can fit both depending on what you measure. The real question is which description captures what the brain is doing when it thinks.
A rough analogy: digital electronics treats crosstalk and ringing as noise to be engineered away. Miller's group asks whether the brain treats its equivalent as signal. The analogy breaks quickly—nobody designed the cortex—but it captures the shift.
What the MIT researchers are actually proposing
First, the genre. "Analog Cognition and Consciousness" is a review: a theoretical synthesis of prior experimental work, much of it from Miller's lab. It reports no new experiment testing the framework.
The argument begins with something well established. Many neurons in higher cortex respond to mixtures of stimuli, rules, and contexts—a property called mixed selectivity. A 2013 Nature study co-authored by Miller found that such neurons in monkey prefrontal cortex carried distributed information about every task-relevant variable, giving downstream readout a computational advantage. That power creates a coordination problem: if one neuron belongs to many overlapping networks, something must decide which network is active right now—far faster than synapses can be rewired.
The review's answer is brain waves. In the authors' account, slower alpha and beta rhythms, arising from circuits that store goals and memories, regulate faster gamma activity tied to incoming sensory information. The authors describe beta activity as an inhibitory stencil laid across the cortex, marking where incoming signals are suppressed and so where sensory processing can happen. Because these waves travel across the cortex, the control is spatial as well as temporal—what the authors call spatiotemporal computing. Synapses store the representations; wave dynamics help select which ones are active.
The paper uses "analog" in the classical sense: computing with continuous physical quantities that directly represent variables, rather than with binary states. Waves suit this because where they meet, they add, cancel, or form new patterns. In the authors' framing, the way task variables map onto wave geometry specifies the "program." Their illustration is a worked example, not a biological measurement.
Underneath sits ephaptic coupling: electric fields produced by collective neural activity feeding back onto nearby neurons' membrane voltages, no synapse required. This part is not speculation. In rat cortical slices, Costas Anastassiou and colleagues found that extracellular fields shifted membrane potentials by less than half a millivolt yet could strongly entrain spiking, especially at slow frequencies. Flavio Fröhlich and David McCormick showed that weak fields at naturally occurring strengths could entrain network activity in slices. In living animals the evidence is inferential: a 2026 Cerebral Cortex analysis by Dimitris Pinotsis and Miller modeled prefrontal recordings and estimated that field-to-neuron influence exceeded the reverse—a statistical result, not a direct manipulation.
So: mixed selectivity is established, traveling waves are well documented, and ephaptic effects are demonstrated in tissue and supported by modeling in vivo. The review builds a theoretical framework from those findings, but the proposed computational role of wave interference still requires direct testing, and Miller is candid about it. "This is a theory," he told MIT's news office, adding that his lab plans to search brain-wave patterns for signatures of analog computation.
Signal, computation, or both?
The hardest problem for any wave-based theory is that neurons generate the waves. A wave pattern accompanying a task might be doing the work—or be the sound of the work being done.
Skeptics exist. NYU neuroscientist György Buzsáki told Quanta Magazine that waves largely reflect the synaptic activity beneath them: "You don't need anything happening in the extracellular space."
Developers know this shape of problem. A CPU's power draw tracks its workload so closely that side-channel attacks can decode a program from it. Nobody concludes the power trace is executing the code.
Separating correlation from causation requires intervention: perturbing fields while holding synaptic activity as constant as possible, then checking whether computation changes as predicted. Slice experiments have done versions of this; doing it cleanly in a behaving animal is far harder. A hypothetical research question: would disrupting a wave's interference pattern, without silencing the neurons generating it, selectively impair the computation the theory assigns to it?
A simulation is not an explanation
Suppose the theory is correct. A digital computer could almost certainly simulate it; field dynamics are standard numerical fare.
But "reproduce" hides at least four achievements: simulating a process, reproducing its functional organization, reproducing its physical mechanism, and reproducing whatever experience accompanies it. A weather simulation models a hurricane without getting anything wet. That analogy is often used as a knockdown argument against machine consciousness, and it isn't one. Whether consciousness is more like wetness or more like arithmetic is precisely what's disputed.
Serious researchers occupy both positions. A 2023 report led by Patrick Butlin and Robert Long assessed AI systems against "indicator properties" drawn from theories of consciousness, adopting computational functionalism as a working hypothesis. Anil Seth, by contrast, argued in Behavioral and Brain Sciences in 2025 that consciousness may depend on our nature as living organisms.
The MIT theory doesn't settle this; it sharpens it. If part of the computation is carried by field dynamics, whether a simulation of them counts as "the same computation" becomes less a thought experiment and more an engineering spec.
From cognition to consciousness
The review's boldest step is from flexible thought to awareness. Its abstract hedges; its conclusion is firmer, proposing that consciousness emerges when wave dynamics carrying the brain's predictive models organize the cortex into a globally integrated state. Key support comes from anesthesia. In a 2025 Cell Reports study of two monkeys, Miller's lab and anesthesiologist Emery Brown found that ketamine and dexmedetomidine, drugs with very different molecular actions, produced similar shifts in prefrontal phase alignment.
That is evidence about the transition between conscious and unconscious states, not an explanation of why any state feels like something. A marker is not a mechanism, and a mechanism of coordination is not yet an account of experience. That gap—the hard problem—is a philosophical challenge, not a scientific result. It doesn't make investigation impossible; it makes it slow.
The theory enters a crowded room. Global Neuronal Workspace theory (GNWT) holds that information becomes conscious when it ignites and is broadcast across cortical networks. Integrated Information Theory (IIT) ties consciousness to a system's intrinsic cause-effect structure. Neither is settled. In a preregistered adversarial collaboration published in Nature in 2025, IIT's predicted sustained synchronization in posterior cortex did not appear, and neither did GNWT's predicted ignition when a stimulus disappears. A large group of researchers has also argued in Nature Neuroscience that IIT's core claims are untestable in principle, a charge that remains hotly contested.
The MIT authors place themselves in partial agreement with both. Like GNWT, they see consciousness in large-scale cortical integration, but they locate that integration in field interactions rather than broadcasting. Like IIT, they stress organization and self-influence, but they place the mechanism in cortex-wide wave dynamics rather than a posterior "hot zone." That is a real mechanistic disagreement, and a useful one: whether waves replace broadcasting or implement it is something experiments could eventually decide.
Debugging consciousness with the wrong abstraction
Every engineer has met the bug that lives below the abstraction: the race condition that exists only because of timing. The model was right about the logic and wrong about the system.
The opposite mistake is just as common. Good abstractions exist to ignore details that don't matter; an engineer who trusts none of them never ships. The skill is knowing which omitted details matter for the question you're asking.
That's the most useful way to read the MIT proposal. A spike-and-synapse model may predict behavior while missing how the brain coordinates itself in real time; a wave model may explain flexible cognition without explaining experience.
For AI, the lesson cuts several ways. Large language models show that impressive capabilities can emerge from architectures sharing almost nothing with cortical wave dynamics, so capability and biological resemblance can come apart. That says nothing direct about experience; the MIT theory shows neither that such systems are conscious nor that they cannot be. Analog or neuromorphic hardware wouldn't settle it either, because physical resemblance is not the same as having the right properties, whatever those turn out to be.
What would change our minds
The theory makes commitments that can fail. The review predicts, for instance, that prefrontal neurons should be spatially organized by goals and contexts, and that shifts between representations should appear as beta activity moving from one spatial pattern to another. Confirming those would help; targeted field perturbations that alter cognition as predicted would help more. If waves prove to be a faithful readout of synaptic computation with no independent causal role, the analog layer collapses back into a measurement tool.
Even a demonstrated causal role in cognition wouldn't automatically extend to consciousness, and the review offers no predictions specific to experience. Those would need to show that the proposed wave organization tracks experience itself, not just responsiveness, wakefulness, or the capacity to report.
We have mapped the connections, recorded the signals, and built networks that write and debug code. What we haven't settled is which properties of biological computation do the work—and which, if any, make it feel like something to be the machine.
If intelligence can be reproduced in software but experience depends on the physics doing the computing, a system might pass every behavioral test we give it. And we still wouldn't know what, if anything, it experiences.
Sources:
Cell Reports article page https://www.cell.com/cell-reports/fulltext/S2211-1247(25)00456-5
ScienceDirect record https://www.sciencedirect.com/science/article/pii/S2211124725004565
Top comments (2)
What I valued most is how carefully you kept the evidence tiers apart. Mixed selectivity is established, traveling waves are well documented, ephaptic effects are demonstrated in slices, and the computational role of wave interference is still a theory. That ladder is exactly what a topic this seductive needs. Your CPU power-draw analogy also captures the central difficulty better than most neuroscience write-ups do: a signal that tracks the computation so closely is not necessarily performing it. Buzsáki’s skepticism and the proposed intervention test, disrupting the interference pattern without silencing the neurons that generate it, show what a falsifiable version of this claim would look like. It is rare to see an article on consciousness say plainly that a marker is not a mechanism. Thank you for writing something this rigorous and readable.
The distinction between simulating a process, reproducing its functional organization, reproducing its physical mechanism, and reproducing the experience that accompanies it is the part I will be thinking about longest. The weather-simulation analogy is usually wielded as a knockdown argument, and you rightly note that it presupposes the very thing in dispute: whether consciousness is more like wetness or more like arithmetic. I also liked the framing that if field dynamics carry part of the computation, the question of whether a simulation counts as “the same computation” becomes less a thought experiment and more an engineering spec. Your closing image, a system that passes every behavioral test while we still cannot say what, if anything, it experiences, is a sobering place to land for anyone working on AI. A question for you: do you think the Butlin and Long indicator-property approach would need to change if wave dynamics turn out to be causally essential, or would it simply become another property to check for?