Integrated Information Theory: Putting a Number on Consciousness

The theory begins with the neuroscientist Giulio Tononi, whose foundational paper appeared in 2004. Rather than start from the brain and work toward experience, IIT starts from experience itself and asks what any physical system would need in order to have it. In its mature form, laid out in 2014 and refined since, it distills experience into five axioms, claims so basic they are meant to be undeniable: your experience exists for you, from the inside; it is structured, made of parts; it is specific, this scene and not another; it is unified, not a bundle of separate pieces; and it has definite borders, a clear edge where it stops.

From those axioms IIT derives Phi, its measure of integrated information. The intuition is that a conscious system must be more than the sum of its parts: the whole must generate information that no independent collection of the pieces could. To compute Phi, you carve the system every possible way and find the weakest cut, the partition that loses the least when severed. Phi is how much is still lost even across that weakest cut. High Phi means the system genuinely cannot be reduced to independent parts. That irreducibility, IIT claims, is what experience is made of.

The sharpest prediction concerns machines. A feed-forward network, one where information flows in a single direction through its layers without looping back, has a Phi of exactly zero, however complex it is. Because today's leading artificial intelligence, including the large language models behind the current boom, is built on feed-forward architectures, IIT's verdict is blunt: these systems, whatever they say, are not conscious. Note carefully what is Tier 1 here. That this is what the theory predicts is a plain fact about the math. Whether the prediction is true of the world is a further question, and a far more uncertain one.

Taken as a claim about reality rather than about the theory, the consequence is radical. If IIT is correct, no feed-forward system could ever be conscious no matter how human its behavior, while a relatively simple circuit with the right looping connectivity could have genuine experience. Stranger still, IIT holds that a perfect digital simulation of a brain would feel nothing, because a computer running through sequences of simple gates lacks the intrinsic, woven causal structure that Phi measures. That directly contradicts the functionalist view that consciousness rides on what a system computes, not what it is made of. It is IIT's most controversial practical claim, and it is unproven.

The theory also reaches in the other direction, toward the very small. Any system with a Phi above zero has some flicker of experience. A single photodiode, on IIT's accounting, has a tiny but nonzero Phi, and therefore an almost unimaginably simple shred of what it is like to be it. Tononi accepts this openly: IIT implies a form of panpsychism, though a graded and quantitative one, a smooth scale from the barely-there experience of a diode up to the rich inner life of a brain, not the claim that rocks daydream. Whether that implication is a strength or a fatal embarrassment is one of the fault lines of the whole debate.

Closer to home, IIT points at a specific piece of anatomy. It predicts that consciousness lives in the posterior cortex, the parietal and occipital regions at the back of the brain, rather than the prefrontal cortex at the front. The evidence fits: damage to the prefrontal cortex does not extinguish consciousness, while damage to the posterior hot zone can erase specific contents of experience. This has become one of IIT's genuine points in its favor, and it sets up a direct collision with its main rival, Global Workspace Theory, which bets on the front of the brain.

The tool is called the Perturbational Complexity Index, or PCI, introduced by Adenauer Casali and colleagues in 2013. It works exactly the way IIT would suggest a consciousness meter should. Send a magnetic pulse into the cortex, record the electrical echo with EEG, and measure how complex that echo is. Across 208 measurements spanning wakefulness, sleep, anesthesia, coma, the vegetative state, and locked-in syndrome, PCI sorted conscious from unconscious states with better than 95 percent accuracy. It is the closest thing the field has to an objective consciousness detector.

The numbers line up cleanly with awareness. Waking and dreaming both score high, around 0.4 to 0.7, because both are richly integrated states. Deep dreamless sleep and general anesthesia score low, down around 0.1 to 0.3, as the cortex fragments into local, disconnected responses. Most striking of all, PCI has flagged high complexity in some patients diagnosed as vegetative, hinting at a conscious mind trapped in an unresponsive body, and it correctly identifies locked-in patients, fully aware but paralyzed, as conscious. That last result is not just a scientific win but an ethical one, with direct consequences for how such patients are treated. This is IIT's strongest empirical achievement, and even critics grant it.

One honest caution has to travel with that success, because it is routinely overstated. PCI does not measure Phi. Nobody has ever computed Phi, in its full mathematical sense, for a human brain, and nobody can with current methods. PCI is inspired by IIT and measures a related kind of complexity, but the theory's actual quantity remains out of reach. The clinical tool works; that is not the same as the theory being proven.

Computing Phi means checking every possible way to partition a system, and the number of partitions explodes super-exponentially as the system grows. William Marshall showed in 2016 that even approximation methods hit this wall. The scale is hard to overstate: a system of only a few hundred simple on-or-off elements already has more possible partitions than there are atoms in the observable universe. A real brain has roughly 86 billion neurons. Calculating its true Phi is not merely hard; it is, by any known method, impossible. IIT's central quantity cannot be measured in the one system we most want to measure it in.

Defenders of the theory reply that intractability is not the same as unfalsifiability. Plenty of real quantities in physics cannot be computed exactly, and proxies like PCI let the theory make contact with data anyway. Critics counter that a theory whose defining measure can never be checked against its own definition, in any real brain, is in a strange scientific position no matter how elegant the mathematics. Both are fair points, and the gap between them is where much of the fight lives.

Not every consequence that sounds absurd actually counts against the theory, and honesty cuts both ways here. IIT predicts that the cerebellum, which holds about 69 billion neurons, roughly 80 percent of all the neurons in the brain, contributes almost nothing to consciousness, because its circuitry is largely feed-forward and modular rather than richly interconnected. That sounds wrong until you check the clinic: people born without a cerebellum, or who lose it, keep their conscious experience largely intact, losing coordination rather than awareness. Here the counterintuitive prediction is actually correct, which is a real point in the theory's favor even as other predictions cut against it. And the philosopher Philip Goff defends the panpsychist implication outright, treating IIT as the most rigorous framework yet for taking consciousness to be fundamental.

That September, an open letter signed by 124 consciousness researchers used a word rarely thrown in public science: pseudoscience. The signatories argued that IIT's empirical support was being oversold and that its panpsychist implications put it outside real science. Tononi and colleagues fired back that the letter confused being untestable in practice, which they conceded, with being unfalsifiable in principle, which they denied, and pointed to PCI as genuine empirical progress. The exchange was unusually bitter for a scientific disagreement.

The letter is also a lesson in reading past a headline. A follow-up survey of the broader consciousness-research community, not just the letter's signatories, found that among the 60 who responded, only about 8 percent fully agreed with the pseudoscience label, while about 20 percent rejected it outright and most landed somewhere in between. Many who signed objected specifically to panpsychism, not to IIT's mathematics or its clinical predictions. When the field's own biggest experiment reported its results, a Nature editorial went out of its way to say that pseudoscience language has no place in a collaborative scientific process. The strong word made news; the field's actual center of gravity was far more measured.

Beginning in 2019, a Templeton-funded adversarial collaboration set IIT against Global Workspace Theory, with both camps committing in advance to what each theory predicted, then testing it on the same 256 people using fMRI, MEG, and direct intracranial recordings. The results, published in Nature in 2025, were genuinely mixed and do not hand either side a clean win. On the big question of where consciousness lives, the data favored IIT: awareness tracked the posterior cortex, and the frontal ignition that Global Workspace Theory predicted did not show up as required. But on one of IIT's own specific predictions, that the posterior hot zone would show sustained synchronized activity in the gamma frequency band, the data did not cooperate; that signature was not found. IIT came out ahead on the broad question and stumbled on a specific one of its own.

That is the honest shape of it, and even the collaborators disagreed publicly about what it meant. A contest designed to be decisive turned out, like most real science, to be a partial result that moved the field without settling it. IIT looks stronger than its loudest critics allow and weaker than its most fervent boosters claim. Both of those can be true at once, and here they are.