The core idea
Most theories of consciousness start from the brain and ask which activity is conscious. Giulio Tononi started from experience itself. Every experience, he argued, is integrated (you see a red ball, not a separate red and a separate round) and informative (it is one specific experience out of a vast number possible). A physical system can support consciousness only if it has the same properties.
From this, IIT (BMC Neuroscience, 2004; latest version IIT 4.0, Albantakis et al., PLOS Computational Biology, 2023) defines consciousness as integrated information and measures it with Φ.
What Φ measures
Take a system of interacting parts. Ask how much its current state constrains its past and future states as a whole, and compare that with what you would get if you cut it into independent pieces along the weakest seam. The difference is Φ.
- A pile of sand has almost no Φ: cutting it changes nothing.
- A feed-forward circuit, where signals flow one way, has zero Φ however complex it is.
- A densely recurrent network, where parts constantly influence each other, can have high Φ.
IIT also says which part of a system is conscious: the subset with maximal Φ, called the main complex.
Predictions
- Posterior cortex. IIT predicts that the core of experience lies in densely interconnected posterior regions of the cortex, a "posterior hot zone", more than in the frontal areas that report and plan.
- Loss of consciousness. In deep sleep and anaesthesia, integration should break down even if activity continues. The perturbational complexity index (Casali et al., Science Translational Medicine, 2013) tests this: a magnetic pulse to the cortex produces a complex, widespread echo in waking and dreaming, and a simple local one in dreamless sleep and anaesthesia.
- Machines. Because Φ depends on physical causal structure, a conventional computer running any program has very low Φ.
The tests against global workspace theory
A consortium of labs (Cogitate) designed preregistered experiments in which IIT and global neuronal workspace theory predict different outcomes. The results, published in Nature in 2025, supported some predictions of each and contradicted others: the sustained synchronisation within posterior cortex that IIT predicted did not appear. The exercise changed the field more than the verdict did: theories of consciousness now compete on predictions that can fail.
Criticism
- Computability. Exact Φ requires checking every way of cutting a system, which is impossible for a brain.
- Counterintuitive cases. Scott Aaronson showed in 2014 that simple grid-like arrangements of logic gates can have Φ far higher than a human brain, which many take as a reductio.
- Testability. In 2023 more than a hundred researchers called IIT pseudoscience in an open letter; others replied that it makes more testable predictions than most rivals. The disagreement is still live.
Alternatives to Φ
Several groups propose other integration measures that are cheaper to compute. The Pointer Architecture preprint on this site defines a substrate-side measure, Φ_PA, which is non-zero only for an observer that reads its own state back, with non-trivial scope and lifespan. It predicts, for example, zero for a single transformer forward pass and a positive value when the model reuses its own cache. The predictions are so far checked on toy encodings. Details are on the preprint page.
Bottom line
IIT is the most mathematically developed theory of consciousness. It explains why integration matters and has produced a useful clinical tool. Its central quantity cannot be computed for real brains, and some of its implications strike many researchers as absurd. It is a serious theory that is still under active test.