Integrated Information Theory: A Mathematical Theory of Consciousness

The Core Idea: Consciousness as Integrated Information

IIT starts from a simple observation: conscious experience has certain properties that any theory of consciousness must account for. Your experience right now is specific (it is this experience and not some other), unified (it is a single, integrated experience rather than a collection of separate parts), definite (it has a particular structure with clear boundaries), and informative (it rules out a vast number of other possible experiences). IIT calls these the axioms of consciousness, and from them derives postulates about what physical properties a system must have to generate conscious experience.

The central measure in IIT is phi, which quantifies how much integrated information a system generates. Integrated information is the amount of information generated by a system as a whole that exceeds the information generated by its parts independently. A system with high phi is one whose components interact in a way that creates information that cannot be reduced to the sum of its parts.

Consider a simple example. A photodiode has two states (on or off) and generates one bit of information, but it is not integrated because it has no internal structure, no parts that interact. Your brain, by contrast, has billions of neurons whose interactions create enormously more information than those neurons would generate independently. According to IIT, the difference in phi between the photodiode and the brain is the difference in their consciousness.

The Five Axioms and Postulates

IIT is built on five axioms, properties of consciousness that are taken as self-evident from the first-person perspective, and five corresponding postulates, requirements that any physical substrate of consciousness must satisfy.

Intrinsic existence: Consciousness exists from its own intrinsic perspective, independent of external observers. A conscious experience exists for itself, not merely as an abstract description.

Composition: Consciousness is structured. An experience is composed of many distinguishable elements, such as colors, shapes, sounds, and thoughts, that combine in specific ways.

Information: Each experience is specific. It is this particular experience, different from all other possible experiences. This specificity is what makes experience informative.

Integration: Consciousness is unified. An experience cannot be decomposed into independent sub-experiences. You do not have separate, independent experiences of color and shape; you have one integrated experience that includes both.

Exclusion: Consciousness has definite boundaries in both content and spatiotemporal grain. At any given moment, you have one experience with a specific content, not multiple overlapping experiences at different levels of detail.

What IIT Predicts About AI

IIT makes a striking and controversial prediction about digital computers: they have very low phi regardless of what software they run. This is because standard computer architectures process information through logic gates that operate in a largely feedforward fashion, with each gate performing a simple Boolean operation. Even if a computer perfectly simulates a brain, neuron by neuron, IIT predicts that the simulation would not be conscious because the causal structure of the computer (transistors executing instructions sequentially) is fundamentally different from the causal structure of the brain (neurons interacting in a dense, recurrent network).

This prediction is called the "unfolding" argument: a digital simulation "unfolds" the rich, integrated causal structure of the brain into a long sequence of simple operations, destroying the integration that IIT identifies with consciousness. The simulation might behave identically to the brain from the outside, but it would not be conscious on the inside.

If IIT is correct, achieving machine consciousness would require building systems with high intrinsic integrated information, likely using architectures very different from conventional digital computers. Neuromorphic computing, which uses hardware that more closely mimics the parallel, recurrent structure of biological neural networks, might be a more promising substrate for conscious machines under IIT.

Strengths and Criticisms

IIT has several notable strengths. It is mathematically precise, which allows its predictions to be stated clearly and, in principle, tested. It explains why consciousness comes in degrees (systems can have more or less phi). It accounts for the specific quality of different conscious experiences (different phi structures correspond to different qualia). And it makes concrete predictions about which brain states are conscious and which are not, some of which have been confirmed experimentally.

However, IIT also faces significant criticisms. The most practical objection is that calculating phi for any realistically complex system is computationally intractable. The calculation requires evaluating all possible ways of partitioning a system, which grows super-exponentially with the number of elements. This means that IIT predictions about specific systems, whether brains or computers, often cannot be directly tested.

A deeper philosophical objection, sometimes called the "small phi" problem, is that IIT attributes some degree of consciousness to very simple systems, including thermostats and individual logic gates, as long as they have non-zero phi. Many researchers find this implausible and argue that it represents a reductio ad absurdum of the theory, though proponents of IIT embrace this consequence and argue that consciousness is indeed far more widespread than commonly assumed, a position related to panpsychism.

Scott Aaronson, a computer scientist, has raised concerns that IIT would attribute high phi to simple systems that are clearly not conscious, such as large grids of XOR gates. Tononi has responded with modifications to the theory, but the debate highlights the difficulty of connecting mathematical formalism to phenomenological reality.

The Mathematics of Phi

Understanding phi requires grasping several concepts from information theory. At the most basic level, information is measured in bits and quantifies how much uncertainty is reduced when you learn the state of a system. A coin flip carries one bit of information because it resolves one binary uncertainty. Your brain at any given moment is in one of an astronomically large number of possible states, so learning which specific state it is in provides an enormous amount of information.

Integrated information goes beyond simple information by measuring how much the information generated by the whole system exceeds what would be generated by its parts operating independently. To calculate phi, you hypothetically partition the system in every possible way and find the partition that makes the least difference, the minimum information partition (MIP). Phi is the amount of information lost across this minimum partition. If no partition reduces the information at all, phi is zero and the system is not integrated.

The geometry of integrated information, what IIT calls the conceptual structure, determines the specific quality of conscious experience. Each concept (a set of elements in a specific state) contributes a point in a high-dimensional space called qualia space. The shape of this structure is what it is like to be the system at that moment. Two systems with the same conceptual structure have the same experience, regardless of what they are made of.

IIT and the Adversarial Collaborations

In one of the most important developments in consciousness science, the Templeton Foundation funded a series of adversarial collaborations pitting IIT against Global Workspace Theory. The first round of results, published in 2023, tested predictions of both theories using neural recordings from humans performing visual tasks. The results provided support for some predictions of both theories while disconfirming others, but neither theory was decisively refuted.

For IIT specifically, the experiments tested whether sustained neural activity in posterior cortex (as IIT predicts) or a wave of activity reaching frontal cortex (as GWT predicts) accompanies conscious perception. The results favored posterior cortex involvement, supporting IIT, but the story is more complex than either theory initially predicted. Further rounds of testing are ongoing, and the field is moving toward increasingly precise experimental designs that can distinguish between theories.

Regardless of the outcome, these adversarial collaborations represent a major step forward for consciousness science. For the first time, consciousness theories are being treated as genuine scientific hypotheses subject to empirical testing rather than as purely philosophical positions immune to refutation.

For anyone interested in consciousness science, IIT is essential reading. Whether or not it ultimately proves correct, it has set the standard for what a rigorous scientific theory of consciousness should look like, and it has forced the field to grapple with deep questions about the relationship between physical structure and subjective experience that had previously been left vague.