Integrated Information Theory, From 3.0 to 4.0: The Theory That Rules Out Conscious Computers

Abstract

Integrated Information Theory is the only major account of consciousness that attempts to say how much of it a system has, expressed as a number. It is also the only one from which it follows directly that a digital computer running any program whatsoever would experience nothing.

This article follows the theory's own formulations — version 3.0 published in 2014 and version 4.0 in 2023, both in PLOS Computational Biology — to set out its structure, its most consequential implication, and the reasons it remains contested.

We approach it as a piece of reasoning to be understood rather than a claim to be accepted or dismissed, because its logic is unusual and most summaries get it backwards.

1.It Starts From Experience, Not From the Brain

Almost every other approach begins with the brain and asks which activity accompanies consciousness. IIT inverts this. It begins with properties of experience that it takes to be immediate and irrefutably true of every conceivable experience, and then asks what a physical system would have to be like to have them.

Five such properties are proposed, and the theory calls them axioms of phenomenal existence. Every experience is for the experiencer — intrinsicality. It is specific — information. It is unitary — integration. It is definite — exclusion. And it is structured — composition.

Each axiom is then translated into a requirement on physical systems, which the theory calls a postulate. The distinction is made explicitly: phenomenal existence is immediate and irrefutable, an axiom, while physical existence is an explanatory construct, a postulate, assessed operationally.

That inversion is the source of both the theory's appeal and the objections to it. It takes the thing everyone else finds hardest to explain and treats it as the starting premise.

2.What the Physical Requirement Amounts To

The theory's criterion for physical existence is causal: to be is to have cause-effect power. A system exists physically to the extent that its parts can take and make a difference to one another in a reliable way, ideally captured by its full transition probability matrix.

The substrate of consciousness must then satisfy all five properties in physical terms. Its cause-effect power must be for itself, specific, unitary, definite and structured.

From the composition requirement comes the mathematics. Subsets of units specify cause-effect states over subsets of units — distinctions — which can overlap with one another — relations. Together these form what the theory calls a Φ-structure, and the sum of its component values gives Φ, which the theory identifies with the quantity of consciousness.

Version 4.0's advances are technical rather than conceptual: a more precise expression of the axioms as postulates, a new measure of intrinsic difference constructed to be consistent with the postulates, and explicit assessment of causal relations.

3.The Implication That Divides the Field

Here is the consequence that makes this theory matter for anyone thinking about artificial minds, stated in the 4.0 paper itself.

Digital computers implementing programs capable of artificial general intelligence may in principle be able to emulate any function performed by conscious humans — and yet, because of the way they are physically organised, would do so without experiencing anything, or at least anything resembling what each of us experiences in quantity and quality.

The reasoning follows from the framework rather than from any intuition about machines. What matters on this account is the intrinsic cause-effect structure of the physical substrate, not the function it computes. A system can reproduce every input-output relation of a conscious being while having a causal architecture that supports almost no integrated information.

This is a genuinely substantive prediction and it is the reverse of what Global Workspace Theory implies. If consciousness is a broadcasting function, then building the architecture might suffice. If it is a property of causal structure, then a functionally identical system built differently would be dark inside.

4.Why It Is Contested

Three objections recur, and none of them is a claim that the theory is obviously wrong.

The first is testability. If Φ is difficult to compute for anything more complex than a small network, and if the theory predicts experience in systems that cannot report it, then the routes by which it could be falsified are narrow. The adversarial collaboration described elsewhere in this library is the most serious attempt to construct one, and its synchrony findings did not support the theory's predictions.

The second is where the axioms come from. Calling a property of experience immediate and irrefutable is a strong move, and a reader who does not share the intuition has no obvious way to argue about it.

The third concerns the implications. Because Φ is a property of causal structure rather than of biology, simple systems with the right architecture would have some non-zero amount of experience. Whether that is a bold prediction or a reductio depends on prior commitments the evidence does not settle.

5.How We Read It

IIT is the most formally developed theory in this area and the hardest to test. It offers something no rival does — a principled account of why a physical system would have experience at all rather than merely processing information — at the cost of resting on premises that must be granted before the argument starts.

For readers of writing about artificial consciousness, one implication is worth carrying: on this account, passing every behavioural test would establish nothing. Any argument that a system is conscious because of what it can do is an argument this theory rejects at the foundations.

Editorial Comment

MindHeaven® makes no claim relating to consciousness, artificial or otherwise, and nothing here concerns any product.

It is worth noting that this theory, taken at face value, would rule out our own AI Consciousness Lab ever producing a conscious system on conventional hardware. We publish it as it stands rather than as we might prefer it, which is the only way a library of this kind is worth anything.