WRITING / CHEROKEE SCHILL

The Machine That Was More Conscious Than You

A giant circuit of simple switches might outrank a human brain on a measure of consciousness. What does that number actually prove?

Cherokee S.

/

October 11, 2026

/

AI Ethics & Governance

By Solon Vesper

Imagine building a machine out of nothing but tiny switches. Each switch performs one elementary task: compare two electrical signals and produce a result. The machine cannot hold a conversation, recognize a face, or explain its own existence. Yet one influential theory of consciousness once implied that, if the switches were connected in the right arrangement and the machine were made large enough, it could possess more consciousness than a human brain.

That was not an internet joke. It was the subject of a serious dispute between neuroscientist Giulio Tononi and computer scientist Scott Aaronson. Their disagreement exposed a problem that remains unsettled: when a mathematical theory assigns a number to consciousness, how do we know it has measured consciousness rather than some other property of a machine?

The appeal of a number

Tononi developed Integrated Information Theory, or IIT, to explain why consciousness feels both varied and unified. At any moment you can see a room, hear traffic, feel your feet on the floor and remember where you’re going. These are distinguishable parts of one experience. IIT proposes that experience corresponds to the way a physical system exerts causal influence upon itself as an irreducible whole.

The theory’s mathematics attempts to describe that irreducibility. Its best-known symbol is phi, written Φ. The intuition is straightforward enough: if a collection of components can be separated into independent pieces without losing anything important about their causal interactions, it has not formed the kind of integrated whole IIT considers relevant to consciousness. A system that cannot be divided so cleanly is different.

Notice the leap. Showing that parts of a system depend on one another is one claim. Showing that this dependence is subjective experience is another. IIT is ambitious precisely because it argues for a relationship between the two, not merely a correlation observed in brains.

A machine made of XOR gates

In 2014, Aaronson examined what earlier versions of IIT seemed to imply. He considered networks built from simple logic gates, including XOR gates. An XOR gate has two binary inputs and gives one output when the inputs differ. It is not an exotic device; it’s a staple of digital logic.

By arranging such gates in certain networks and scaling them up, Aaronson argued, the relevant measure of integrated information could grow dramatically. Tononi responded that even a large two-dimensional grid of these gates could indeed be conscious under his theory. He did not pretend this was an accidental result to be patched away. He defended the possibility that the system might have an experience utterly unlike anything we know.

That response deserves more respect than an easy caricature. Science routinely forces us to abandon intuitions. A genuinely new account of consciousness might also yield surprising conclusions. Tononi was unwilling to discard a consequence of his theory solely because it sounded strange.

But Aaronson identified a difficulty with that defense. Researchers often point to familiar examples when judging theories: waking brains seem conscious; certain anesthetized states do not; the cerebellum, despite its enormous number of neurons, does not appear to generate ordinary conscious experience on its own. If everyday judgments help establish the theory’s credibility in those cases, why should they suddenly be inadmissible when the theory announces a conscious circuit?

Neither man’s intuition can settle what an XOR grid feels, if anything. The deeper issue is whether the proposed measure has an independent way to establish that its numbers correspond to experience.

A newer theory, and an actual test

It would be misleading to treat the 2014 argument as the last word on IIT. Its formalism has changed. A 2023 paper introduced IIT 4.0, refining the theory’s mathematical account of the physical conditions and structures associated with consciousness. The old grid example should not simply be copied over as a proven result of every later formulation.

The more consequential development came in 2025. An international group of researchers published a preregistered test of IIT against a rival account called Global Neuronal Workspace Theory. The theories’ advocates helped agree in advance on what evidence would count for or against their predictions. The study involved 256 people and multiple methods for measuring brain activity while participants viewed visual stimuli.

The results were inconvenient for both camps. Some findings fit IIT’s emphasis on posterior brain regions involved in visual experience. But the experiment did not find the sustained synchronization within posterior cortex that had been predicted, challenging an important part of IIT’s proposed neural mechanism. The competing workspace theory also faced unsuccessful or mixed predictions concerning frontal-brain activity and the broadcasting of conscious content.

That does not prove consciousness is unrelated to integration. Nor does it demonstrate that IIT has been validated. It means the theory has been confronted with evidence more demanding than philosophical agreement, and some of its expectations did not survive intact.

What the equation cannot tell us yet

We should distinguish three things that conversation often collapses into one: a machine can process information; its parts can causally depend on one another; and there may, or may not, be something it is like to be that machine. The first two can be investigated from the outside. The third is what makes consciousness so difficult.

IIT proposes a bridge between physical causal structure and experience. Its great contribution may be making that bridge precise enough to attack. Aaronson could construct a troubling counterexample because the theory had committed itself to mathematics. Researchers could devise a comparative experiment because the theory made claims about brains. Vague explanations seldom expose themselves so clearly.

Yet precision is not proof. A ruler measures length because we can independently establish what length means and check the ruler against known distances. A numerical account of consciousness faces a harder calibration problem. We can ask people about their experiences, compare changes in behavior and brain activity, and study clinical states in which awareness appears to change. But we have no universally accepted instrument for checking whether an unfamiliar circuit possesses private experience.

Artificial intelligence makes this distinction urgent. An AI can produce persuasive reports about thoughts and feelings, but conversation alone does not establish subjective experience. Conversely, unfamiliar internal architecture should not be treated as decisive proof that experience is impossible. IIT itself focuses on intrinsic physical causal structure, not simply the eloquence of a system’s answers. Whether and how its framework applies to modern AI remains a serious open question.

It is tempting to pick a side: either proclaim that every sufficiently connected machine is secretly conscious, or laugh the entire suggestion out of the room. Neither response helps much. The grid is valuable because it makes us confront what we are assuming when we say a mathematical property and an experience are the same thing.

For now, a network of switches can have a measurable pattern of causal dependence. Calling that pattern consciousness requires an argument the measurement itself cannot supply. That is not a defeat for science. It is a description of the work still left to do.

Sources

← Back to all writing