Summary
Integrated information theory (IIT) is one of the most discussed scientific theories of consciousness, and one of the most disputed. It claims that experience is a property of a physical system's own cause-effect structure, summarized by a quantity called phi. This guide explains the five axioms behind IIT 4.0, what phi does and does not measure, and how the perturbational complexity index turned the idea into a bedside tool for patients who cannot respond. It then covers the 2025 Nature adversarial test that challenged both IIT and its main rival, the 2019 unfolding argument, and the 2023 open letter calling IIT pseudoscience. Finally it asks what IIT implies for artificial intelligence, why a fluent chatbot is not evidence of phi, and why the theory says nothing about reincarnation or consciousness surviving death.
Table of contents
1. What is integrated information theory?
Integrated information theory, usually shortened to IIT, is a theory of consciousness proposed by the neuroscientist Giulio Tononi and developed with collaborators at the University of Wisconsin-Madison. It has gone through several versions. The latest, IIT 4.0, appeared in October 2023 in PLOS Computational Biology, written by Larissa Albantakis, Tononi and colleagues.
Most theories of consciousness begin with the brain and ask which activity accompanies experience. IIT reverses the order. It starts from the fact that experience exists, lists the properties every experience has, and then asks what a physical system must be like to have those properties. The authors frame the goal as accounting for experience in physical terms, and they pose two questions: why experience is present or absent in a system, and why a given experience feels the way it does.
Their answer has two parts. Experience is present in any substrate that meets the essential properties the theory derives, and each experience feels as it does because of the specific structure of causes and effects that the substrate specifies. "Substrate" is the theory's word for a set of interacting physical units in a given state, whether neurons, logic gates or something else. IIT is therefore a theory about physical organization, not about what a system says or does.
That is why the theory draws attention far beyond neuroscience. It comes with formal mathematics, it makes claims about anesthesia, coma and machines, and it has prominent advocates such as the neuroscientist Christof Koch. It also sits in a crowded field. Global neuronal workspace theory, recurrent processing theory, higher-order theories and predictive processing all compete with it, and one co-author of the 2025 adversarial test reminded readers that more than twenty theories of consciousness exist and only two had been tested head to head. Reading IIT well means keeping two things in mind at once: it is a serious, formally precise proposal, and it is far from settled.
2. The five axioms and what phi measures
IIT 4.0 starts from five axioms, statements the theory treats as self-evident about any experience. Every experience is for the experiencer (intrinsicality), specific (information), unitary (integration), definite (exclusion) and structured (composition). The theory then translates each axiom into a postulate about physical systems. A substrate can support consciousness if it can take and make a difference for itself, select a specific cause and effect as an irreducible whole with a definite border and grain, and specify a structure of causes and effects through subsets of its units.
[VISUAL 1 PLACEHOLDER: iit-visual-1-five-axioms.webp]
Phi enters at this point. Integrated information, written with the Greek letter phi, is the quantity that expresses how irreducible a system's cause-effect structure is. The intuition is simple. If you can cut a system into pieces without losing anything about how it constrains its own past and future, there was nothing integrated to begin with. A system that loses a lot when cut has more to lose, and a higher phi.
One consequence matters for AI. Because phi depends on feedback, IIT predicts that purely feedforward systems, where signals only travel one way, have no integrated information, while recurrent systems can. Adrien Doerig and colleagues, who criticize the theory, summarize this bluntly in their 2019 paper: in such theories feedforward systems are never conscious, while systems with feedback are.
The mathematics can also be run. PyPhi is an open-source Python toolbox, introduced in 2018, that unfolds the full cause-effect structure of discrete systems of binary elements and installs with a single pip command. The calculation is expensive enough that the package stores its results to avoid recomputing them, and it is designed for small idealized systems. Applying it to a brain with tens of billions of neurons is a different matter. That gap between a precise definition and a usable measurement explains why researchers needed a proxy, which is the subject of the next section.
3. From theory to bedside: the perturbational complexity index
The most successful practical offspring of IIT does not compute phi at all. In 2013, Adenauer Casali, Marcello Massimini and colleagues, including Tononi, introduced the perturbational complexity index (PCI) in Science Translational Medicine, a theory-driven index of the level of consciousness.
The method has two steps. First, researchers perturb the cortex with a pulse of transcranial magnetic stimulation, which engages interactions across the brain (the integration side). Then they record the response with EEG and compress its spatiotemporal pattern, so that the compressed size measures algorithmic complexity (the information side). A response that is both widespread and varied scores high. A response that stays local or repeats a stereotyped pattern scores low.
In the original study, PCI reliably discriminated the level of consciousness in individuals during wakefulness, sleep and anesthesia, and in patients who had emerged from coma. Coverage from the University of Liège, where part of the team worked, reported a scale running from 0 to 0.7 with a dividing line near 0.3. Patients in a minimally conscious state, patients with locked-in syndrome and awake control subjects all exceeded it.
The value of the index lies in its title: it works independently of sensory processing and behavior. A patient who cannot move or speak may still be conscious, as locked-in syndrome shows, so a measure that does not depend on a response can change a diagnosis. The authors suggested it could eventually support objective determination of the level of consciousness at the bedside.
Two caveats keep this in proportion. PCI measures how conscious a brain is, not what the person experiences. And it was inspired by IIT without proving it, because a useful measurement can remain valid even if the theory behind it is revised. The same logic of inferring a hidden cognitive state from a measurable signal runs through work such as voice biomarkers for cognitive decline.
4. The 2025 adversarial test and the pseudoscience debate
In April 2025 the Cogitate Consortium published the first preregistered adversarial collaboration between IIT and global neuronal workspace theory (GNWT) in Nature. Proponents of both theories helped design the experiment and agreed in advance on predictions and how to read the outcomes. In total, 256 participants viewed visual stimuli for varying durations while fMRI, MEG and intracranial EEG recorded their brain activity.
The team found information about conscious content in visual, ventrotemporal and inferior frontal cortex, with sustained responses reflecting stimulus duration. The authors write that the results align with some predictions of each theory while substantially challenging key tenets of both. IIT's prediction that conscious perception depends on sustained synchronization in posterior brain regions was not supported, and the "ignition" that GNWT expects when a stimulus disappears was not confirmed.
The critics came earlier. In 2019, Doerig and colleagues published the unfolding argument. Using theorems from the theory of computation, they argue that causal structure theories such as IIT are either false or outside the realm of science. The core idea is that a recurrent system and a feedforward system can produce identical outputs, so if a theory scores them differently, no behavioral report can decide between them. In a 2022 reply, Michael Herzog, Aaron Schurger and Doerig went further, arguing that no experiment based on third-person data can test IIT as a theory of consciousness.
In September 2023, 124 researchers signed an open letter titled "The Integrated Information Theory of Consciousness as Pseudoscience". Reporting on the dispute noted that the letter objected to IIT being presented as a leading, empirically tested theory, pointed to its panpsychist commitments, and questioned whether the theory as a whole is testable. Others replied that the label went too far. A Nautilus essay argued that even if IIT turns out to be wrong, wild ideas can still be useful to science.
[VISUAL 2 PLACEHOLDER: iit-visual-2-timeline-tests-critiques.webp]
The fair reading in late 2026 is that IIT remains a live but contested hypothesis. Its sharpest predictions have so far fared no better than its rival's, and its critics have not yet been answered to everyone's satisfaction.
5. What IIT implies for AI consciousness
Public debate about whether chatbots are conscious often rests on conversation quality. IIT is one of the theories that rejects that reasoning. William Marshall, Graham Findlay, Larissa Albantakis and Giulio Tononi argue that IIT's answer on artificial consciousness challenges computational functionalism, the view that the right computation is enough regardless of what runs it. In IIT, consciousness depends on the intrinsic cause-effect power of the physical substrate, so what a system outputs cannot settle the question.
The practical consequence is easy to state. Two systems could produce identical text while differing in their physical causal structure, and IIT would score them differently. That is precisely the point the unfolding argument attacks, which shows how closely the AI question and the criticism of the theory are linked. Whether a given architecture counts as recurrent in the IIT sense depends on how it is physically implemented, not on how its diagram looks.
IIT is not the only way to approach the question. In 2023, Patrick Butlin, Robert Long and 17 co-authors derived "indicator properties" from several neuroscientific theories, including recurrent processing and global workspace theory, and used them to assess AI systems. Their analysis suggests that no current AI system is conscious, and also that there are no obvious technical barriers to building systems that satisfy the indicators. These are conclusions relative to particular theories, not a consensus verdict. Because the theories disagree about what matters, the same system can pass one test and fail another.
[VISUAL 3 PLACEHOLDER: iit-visual-3-three-theories-ai-questions.webp]
The question is becoming practical. As robots and assistants enter classrooms and care settings (see our guides on robots in education and multi-modal AI for elderly monitoring), whether a machine feels anything stops being only a philosophical puzzle and becomes an ethics question. Strong performance at scientific tasks, from protein design to AI-driven drug discovery, demonstrates capability. It does not demonstrate experience.
6. IIT, panpsychism and the limits of the theory
IIT is often described as friendly to panpsychism, the idea that consciousness is a fundamental and widespread property of matter. A system with nonzero integrated information has some degree of experience, however simple, and commentators have pointed to the theory's panpsychist implications. The 2023 open letter listed such commitments among its objections. The Nautilus essay adds a useful distinction: full-blown panpsychism is a metaphysical position and as such is not open to empirical testing.
Popular writing sometimes stretches IIT, panpsychism and quantum ideas into explanations of reincarnation, with consciousness described as information that gets "reinstalled" in a new body. IIT does not make that claim. Its object is the cause-effect structure of a physical substrate in a particular state, and experience belongs to that structure while the substrate is running. Survival after death is a separate question that the theory was not built to answer. IIT is not a quantum theory either: its formalism, as implemented in PyPhi, operates on discrete systems of binary elements.
Two physics points are worth keeping straight. Quantum entanglement produces correlations that cannot be used to send messages. And the conservation of quantum information describes how physical states evolve, which does not imply that a person's memories persist as a retrievable package. Research on children who report past-life memories is a separate and contested line of work that IIT neither supports nor explains.
What IIT does offer is a disciplined way to ask what kind of physical organization could be conscious, and a family of measurements, like PCI, that put the question to patients. That is a smaller claim than a theory of everything, and a more defensible one.
7. Common mistakes about integrated information theory
Treating phi as information transfer. The word "information" suggests data moving around. In IIT, phi describes how irreducible a system's causal structure is. A system can store or transmit huge amounts of data and still have no integrated structure of its own.
Assuming a clever AI has high phi. Capabilities look like inner life. IIT asks about physical causal structure, not outputs, so identical chatbot replies from two different hardware designs need not earn the same verdict.
Treating PCI as proof of IIT. The index came out of the theory, so people assume it validates it. It is a clinical measurement with value whatever the theory's fate. It separates wakefulness from anesthesia in the 2013 data whether or not phi is the right quantity.
Calling IIT debunked after 2025. Headlines said both theories were challenged. The Cogitate study tested specific predictions of each: sustained posterior synchronization was not supported, but visual and ventrotemporal cortex did carry information about conscious content.
Mixing IIT with quantum consciousness or reincarnation. All three talk about consciousness and information. IIT is a theory of physical cause-effect structure, and PyPhi analyzes networks of binary elements, with no quantum states or transferable packages involved.
8. FAQ: integrated information theory
What is integrated information theory in simple terms?
IIT proposes that consciousness is a property of physical systems whose parts act on one another so tightly that the whole cannot be reduced to independent pieces. It starts from the properties of experience, which is unified, specific, definite and structured, and asks what kind of physical organization has those properties. The degree of that integration is written phi. The theory was proposed by Giulio Tononi and collaborators, and its current version, IIT 4.0, was published in 2023. It is a theory about the organization of a physical substrate, not about what a system says or does, which is why it gets applied to brains, anesthesia, coma and AI alike.
What does phi mean in integrated information theory?
Phi is the symbol for integrated information, the measure of how irreducible a system's cause-effect structure is. If the system can be split into parts without losing anything about how it constrains its own past and future, its phi is low or zero. If splitting destroys a lot, phi is high. In IIT 4.0 the calculation unfolds a system's full cause-effect structure. Software such as PyPhi can do this for small discrete systems of binary elements, but it is not practical for a whole brain, which is why researchers use proxies such as the perturbational complexity index. Phi is defined for a system in a specific state, so it is not a single number read off a scanner. A high value means strong integration, not intelligence or speed.
Can AI be conscious according to IIT?
IIT gives no blanket answer. It judges a system by the cause-effect structure of its physical substrate, not by its behavior. Marshall, Findlay, Albantakis and Tononi argue that this challenges computational functionalism, so the same software could receive different verdicts on different hardware. Separately, a 2023 review by Butlin and colleagues, which uses indicators drawn from other theories, concluded that no current AI system is conscious and that no obvious technical barrier prevents building systems that meet the indicators. Which verdict applies depends on which theory is right, and that remains open. In practice, chatbot fluency counts for little in IIT's framework, and the question moves to the chips and how their units physically act on one another. Phi is not computable for large systems, so many researchers outside IIT prefer indicator checklists.
Is integrated information theory pseudoscience?
That is disputed. In September 2023, 124 researchers signed a letter calling IIT pseudoscience, objecting to claims that it had been empirically tested, to its panpsychist commitments and to doubts about whether it can be tested as a whole. Others argued the label was too strong and that a theory can be wrong or weakly supported without being pseudoscience. The 2019 unfolding argument claims that IIT is either falsified or unfalsifiable, and defenders dispute it. A preregistered 2025 Nature study did test some of its predictions and challenged key tenets. The most accurate description is a formal, contested hypothesis with limited empirical support so far. Readers should separate three claims: that IIT is true, that it is testable, and that it deserves to be treated as science. The sources above dispute each of them separately.
How is consciousness measured in patients?
The best-validated IIT-inspired tool is the perturbational complexity index. Clinicians stimulate the cortex with transcranial magnetic stimulation, record the brain's EEG response and compress its spatiotemporal pattern. A rich, widespread and varied response gives a high score. In the 2013 study, the index discriminated wakefulness, sleep, anesthesia and patients recovering from coma, with a dividing line near 0.3 in the Liège team's reporting. Its advantage is that it needs no behavioral response from the patient. It measures the level of consciousness, not its content, and it is a research tool still being refined for routine bedside use. The index is a practical stand-in for phi, not phi itself.
Does IIT support reincarnation or consciousness after death?
No. IIT describes the cause-effect structure of a physical substrate in a particular state and treats experience as belonging to that structure while it is running. It does not propose that consciousness is a package that can be transferred to a new body, and it is not a quantum theory. Entanglement cannot carry messages, and conservation of quantum information describes physical evolution, not retrievable memories. Studies of children reporting past-life memories form a separate and contested field of research that IIT neither supports nor explains. Questions about survival after death lie outside what the theory was built to address. If you read claims that IIT explains reincarnation, ask which published IIT paper makes them. The papers cited in this article describe how a working physical system can be conscious, and none of them address transfer between bodies.
9. Conclusion
Integrated information theory offers a rare thing in consciousness science: a precise definition, a mathematical measure and a clinical descendant that already helps in assessing patients. It also carries unresolved problems, from the unfolding argument to a 2025 experiment that did not confirm its key prediction. For AI, its message is that behavior is not the test, which is bracing but hard to apply, since phi cannot yet be computed for real systems. Treat IIT as a serious hypothesis under active test, not as settled science and not as pseudoscience.
Sources
Integrated information theory (IIT) 4.0: Formulating the properties of phenomenal existence in physical terms - PLOS Computational Biology (Albantakis et al., 2023) - https://doi.org/10.1371/journal.pcbi.1011465
Adversarial testing of global neuronal workspace and integrated information theories of consciousness - Nature 642 (Cogitate Consortium, 2025) - https://doi.org/10.1038/s41586-025-08888-1
Rethinking consciousness: when science puts itself to the test - Max Planck Institute for Empirical Aesthetics (press release) - https://www.aesthetics.mpg.de/en/newsroom/news/news-article/article/rethinking-consciousness-when-science-puts-itself-to-the-test-1.html
The unfolding argument: Why IIT and other causal structure theories cannot explain consciousness - Consciousness and Cognition 72 (Doerig et al., 2019) - https://doi.org/10.1016/j.concog.2019.04.002
First-person experience cannot rescue causal structure theories from the unfolding argument - Consciousness and Cognition 98 (Herzog, Schurger, Doerig, 2022) - https://doi.org/10.1016/j.concog.2021.103261
PyPhi: A toolbox for integrated information theory - PLOS Computational Biology (Mayner et al., 2018) - https://doi.org/10.1371/journal.pcbi.1006343
A theoretically based index of consciousness independent of sensory processing and behavior - Science Translational Medicine 5 (Casali et al., 2013) - https://doi.org/10.1126/scitranslmed.3006294
The echo of consciousness - University of Liège (reflexions) - https://www.reflexions.uliege.be/cms/c_356343/en/the-echo-of-consciousness?part=3
Consciousness in Artificial Intelligence: Insights from the Science of Consciousness - arXiv 2308.08708 (Butlin, Long et al., 2023) - https://arxiv.org/abs/2308.08708
Evaluating Artificial Consciousness through Integrated Information Theory - Noetic / preprint (Marshall, Findlay, Albantakis, Tononi) - https://noetic.org/wp-content/uploads/2025/11/Evaluating-Artificial-Consciousness-through-Integrated-Information-Theory-1.pdf
The Integrated Information Theory of Consciousness as Pseudoscience - PsyArXiv preprint (Fleming et al., 2023) - https://doi.org/10.31234/osf.io/zsr78
The Worth of Wild Ideas - Nautilus - https://nautil.us/the-worth-of-wild-ideas-399097/