From the Phenomenology to the Mechanisms of ConsciousnessIntegrated Information Theory 3.0

For most of the twentieth century, consciousness was considered too soft and too subjective for serious science. Then, Giulio Tononi and colleagues made a move that changed the terms of the problem entirely. Instead of starting with the brain and asking what it does, they started with experience itself and asked what it must be. That inversion — from neuroscience to phenomenology, then back down to physics — is the engine of Integrated Information Theory, or IIT. The problem with the correlational approach is not that it is wrong. It's that it is incomplete. You can map every brain region that lights up during a conscious experience and still be left with the question: why does that physical process produce experience at all, when other processes — equally active and equally complex — do not? Oizumi, Albantakis, and Tononi make this concrete. The cerebellum has more neurons than the cortex, yet it doesn't seem to generate consciousness. A generalized epileptic seizure drives the brain into intense, synchronized firing — and abolishes experience. Deep sleep leaves the brain metabolically active — and awareness disappears. Correlations can document these puzzles, but they can't explain them. IIT's answer is to start from what we actually know with certainty: the structure of experience itself. Oizumi, Albantakis, and Tononi extract that structure as five axioms — existence, composition, information, integration, and exclusion — then translate each axiom into a postulate that constrains how physical systems must be organized to generate experience. Work backward from those postulates and you get, in principle, a way to identify which physical configurations are conscious and which are not. The axioms are worth sitting with, because they do real work. Take information. Every experience is specific — it is what it is by how it differs from every other possible experience. Seeing red rules out blue, green, darkness, and everything else. That specificity means the experience constrains the space of alternatives in a particular way. Integration says experience is unified. You can't decompose the word "SONO" in your visual field into a simultaneous experience of "SO" on the right and "NO" on the left. It's one thing, irreducibly. Seeing a red triangle isn't seeing an uncolored triangle plus a shapeless red patch — the color and shape are bound together. Exclusion closes the trio by insisting that experience has definite borders and a definite grain. At any moment, there is exactly one experience with fully specified content — not a superposition of overlapping partial experiences and not smeared across all spatial or temporal scales. One experience, one resolution, at a time. These three axioms map directly onto the mathematical machinery of IIT 3.0. The central object is the cause-effect repertoire: a mechanism's fingerprint in time. It's the probability distribution over past states that could have produced the mechanism's current state, and the probability distribution over future states the mechanism could produce. Both are computed from the system's transition probability matrix, which specifies how states evolve. Intrinsic information is then the distance between the mechanism's constrained repertoire and the unconstrained one — what the mechanism actually specifies versus what you would expect by chance. Cause-effect information is defined as the minimum of cause information and effect information. A mechanism only counts if it constrains both the past and the future. In the worked OR-gate example with current state ABC equals one hundred, this yields a cause-effect information value of zero point two five for mechanism A. Integrated information — phi — measures irreducibility. You find the partition of the mechanism that, when severed, loses the least information. That's the minimum information partition, or MIP. Phi is the distance between the intact repertoire and the partitioned one across the best such cut. The distance metric throughout is the earth mover's distance: the amount of work needed to rearrange one probability distribution into another. In different examples from the paper, Q-max cause values come out around zero point three three and zero point four four, depending on architecture. The numbers themselves aren't the point — what matters is that phi gives you a precise, calculable measure of how much a system generates as a whole that cannot be recovered from its parts. Now scale up from individual mechanisms to whole systems, and you arrive at the theory's most striking construct: the maximally irreducible conceptual structure, or MICS. A conscious experience isn't a single number. It's a geometrical shape — a constellation of concepts in qualia space, where each axis represents a possible past or future state of the system. For the three-element example in the paper, concept space has sixteen dimensions. Each concept is a point in that space, its location given by its cause-effect repertoire and its size by its Q-max. Integrated conceptual information W — big phi — is the distance between the intact constellation and the best-partitioned version of it. The MICS is the shape that constellation traces; that shape is the quale. W-max is the quantity of experience; the shape is the quality. Exclusion, at the system level, prevents infinite nesting. Only sets of elements that are local maxima of W count as complexes — they can't overlap, and each element belongs to exactly one complex at a time. A recurrently connected subset can form a major complex with high W-max. In the paper's example, a strongly integrated network generates a W-max of zero point seven six with seventeen concepts. Feed-forward elements, with no reciprocal causal connections, cannot form complexes at all. They operate as background conditions or unconscious pathways. Smaller, non-overlapping minor complexes can coexist as paraconscious structures, but they are excluded from and distinct to the dominant complex. This framework generates predictions that are genuinely strange — and genuinely testable. The first is the philosophical zombie prediction. A feed-forward network can implement the same input and output function as a conscious system across many input states. Tononi and colleagues show this explicitly: unfold the recurrent network's memory over time into a chain of nodes and you get the same behavior, but the feed-forward version has no complex and generates no quale. Behaviorally indistinguishable, experientially zero. IIT says these are true zombies — not a thought experiment, but a structural consequence of the theory. The second prediction concerns inactive elements. A set of COPY gates, all in state zero, can still form a complex and specify a MICS. Element A being off can specify an irreducible cause — D had to be off one step back — and an irreducible effect — B will be on one step forward. What matters isn't whether elements are firing. It's whether they are causally structured to fire, embedded in a network with genuine reciprocal influence. The third prediction is about the relationship between complexity and consciousness. Simple systems can be minimally conscious. Tononi and colleagues construct a two-element system — detector D and predictor P — that forms a complex with two concepts and a W-max of one. A photodiode without feedback, detector driving output with no recurrence, is not a complex at all. Meanwhile, highly modular architectures or purely feed-forward networks, no matter how large, remain unconscious. The cerebellum's modularity, not its neuron count, is what IIT says keeps it out of the main complex. These predictions carry immediate stakes. Behavioral and verbal report cannot be the criterion for consciousness — not for infants, not for brain-damaged patients, not for animals, and not for AI systems. Perturbational approaches combining transcranial magnetic stimulation with high-density electroencephalography have been used to test whether information integration breaks down in the ways IIT predicts — across deep sleep, several forms of general anesthesia, and in patients in vegetative or minimally conscious states. The results support the theory's central claim: the loss and recovery of consciousness tracks the breakdown and recovery of integration. Anatomically, IIT points to the corticothalamic system as the likely seat of the dominant main complex — highly differentiated, densely recurrent, and specialized in ways that maximize W-max. There is one honest difficulty Tononi and colleagues don't sidestep. Rigorously computing complexes and their integrated information for large, realistic neural systems is not currently feasible. The calculation scales explosively with system size. That's a real limitation. But it's a frontier, not a refutation. The audacity of the move is worth naming plainly. Most theories of consciousness start with neurons and work upward, hoping to find the place where experience enters. IIT starts with experience and works downward to the physics. It asks: given that every experience is specific, unified, and bounded, what must the physical substrate be doing? The answer is phi — and everything that follows from it. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.