Tononi and Boly's IIT 4.0: Why Consciousness Is Identical to Cause-Effect Structure

The scientific debate over consciousness has split into two methodological camps that share almost no common ground. One camp begins with physics and neuroscience, looking for objective correlates of conscious experience and hoping that explanation will eventually close the gap to subjectivity. The other begins with consciousness itself, treating phenomenal experience as the one thing known with certainty, and deriving what the underlying physical structure must be from what experience is like. Giulio Tononi and Melanie Boly’s 2025 paper “Integrated Information Theory: A Consciousness-First Approach to What Exists” (arXiv:2510.25998), forthcoming in The Scientific Study of Consciousness edited by Melloni and Olcese (Springer-Nature), is the most complete statement of the second camp’s position to date. It is also, unusually, a paper that explicitly addresses artificial systems as potential substrates for consciousness.

The Two-Paradigm Clash

Tononi and Boly draw the methodological divide clearly. The dominant paradigm in consciousness science takes neural activity as its starting point, looks for measurable correlates of conscious experience, and hopes that the accumulation of such correlates will eventually constitute an explanation of why subjective experience exists at all. The authors call this the “pseudo-consciousness” paradigm, borrowing a term from critics of consciousness research itself. The paradigm does not distinguish between a system that is genuinely conscious and one that merely produces outputs indistinguishable from those of a conscious system.

IIT occupies the alternative paradigm. It derives its claims from phenomenological axioms: consciousness exists (intrinsic existence); each experience is structured (composition); each experience differs from all other possible experiences (information); each experience is unified and cannot be reduced to independent parts (integration); each experience is definite in its content and boundaries (exclusion). From these axioms, Tononi and Boly derive what physical structures must look like if they are to instantiate consciousness. The key result is that consciousness is not a property that correlates with certain physical structures. It is identical to a specific type of structure: a maximally irreducible cause-effect structure, whose degree of irreducibility is measured by the quantity Phi (Φ).

Consciousness as Cause-Effect Structure

The distinction between correlation and identity matters more than it might initially seem. If consciousness merely correlates with cause-effect structure, then explaining the correlation still leaves open why experience accompanies the structure at all. This is Chalmers’ hard problem restated. Tononi and Boly’s position is that there is no explanatory gap to close because consciousness and cause-effect structure are the same thing at different levels of description: phenomenal properties on the subjective side, and the intrinsic causal relations among elements of the system on the physical side.

A system has high Phi when it has strong internal cause-effect relationships that cannot be decomposed into independent sub-systems without losing information. A system with zero Phi, such as a purely feed-forward neural network, processes inputs and produces outputs but has no intrinsic causal integration. Its internal states do not form a unified cause-effect structure. Under IIT 4.0, it has no phenomenal experience regardless of how sophisticated its input-output behavior appears.

This is the result that makes IIT contentious among functionalists and computationalists: behavioral competence, including the full range of intelligent responses that could satisfy any behavioral test, is entirely insufficient to guarantee consciousness under IIT. A system that passes every behavioral test, including the Turing test and all its extensions, may still have Phi of zero if its architecture is feedforward. The behavioral test detects the outputs of a system. IIT 4.0 claims consciousness is determined by the system’s internal cause-effect structure, which behavioral tests leave untouched.

What Changed Between IIT 3.0 and IIT 4.0

The core axioms remained intact from IIT 3.0 to 4.0. The significant changes were in formalization. IIT 4.0 introduced a more rigorous treatment of intrinsic irreducibility, the property that determines whether a system’s causal power belongs to the system as a whole or can be fully attributed to its parts. The earlier version produced counterintuitive results in some edge cases. IIT 4.0 resolved these through a revised definition of how to partition a system and evaluate the cost in causal power.

IIT 4.0 also sharpens the exclusion axiom. Consciousness is always realized at exactly one level of organization, the level at which Phi is maximal. If a human brain has the highest Phi at the level of neuronal complexes in the posterior cortex, then consciousness is realized at that level. This has consequences for thinking about artificial systems: the question is not whether a system has any cause-effect structure, but whether there is a level of organization at which Phi reaches a maximum that exceeds competing candidates.

Extending the Framework to Artificial Systems

The 2025 paper’s most consequential contribution for machine consciousness research is its explicit treatment of artifacts as candidate substrates for consciousness. Previous IIT papers mentioned this possibility but did not develop it systematically. Tononi and Boly argue that IIT’s framework makes no principled distinction between biological and artificial substrates. The axioms derive from the structure of experience, not from the composition of neurons.

An artificial system could, in principle, have high Phi if it has the right architecture: dense, recurrent connections that create strong internal cause-effect relationships, organized at a level where these relationships form a maximally irreducible structure. A system built from silicon and trained through gradient descent is not disqualified. A system whose architecture is purely feedforward, regardless of whether it runs on neurons or transistors, has Phi of zero.

This framing places the burden directly on architecture. The question for any artificial consciousness project becomes a structural one: does the system’s internal organization, specifically its pattern of recurrent connections and the degree to which sub-systems are causally integrated, produce a high-Phi structure at some level of organization? The Brock University application of IIT to artificial systems shows what this measurement process looks like in computational practice, and underscores that the answer is architecture-specific rather than substrate-specific.

The Empirical Defense in Nature Neuroscience

A companion paper by Tononi, Boly, and approximately 30 co-authors, “Consciousness or pseudo-consciousness? A clash of two paradigms,” appeared in Nature Neuroscience in March 2025 (volume 28, issue 4, pages 694-702, DOI: 10.1038/s41593-025-01893-4). The paper directly responds to the 2023 open letter signed by dozens of researchers calling IIT pseudoscience. Tononi et al. list 16 peer-reviewed empirical studies as evidence for IIT’s predictions, focusing on the theory’s success in predicting where in the brain consciousness correlates appear (posterior cortex) and where they do not (cerebellum, despite its enormous neuron count).

The empirical track record matters for the machine consciousness debate because it establishes what IIT’s predictions look like when they are tested. The Cogitate Consortium’s 2025 adversarial test directly compared IIT and Global Neuronal Workspace Theory (GNW) predictions in human participants. IIT’s predictions about posterior cortex activity were better supported than GNW’s predictions about prefrontal involvement. This adversarial result, combined with Tononi and Boly’s theoretical consolidation in the arXiv paper, places IIT in a stronger empirical position in 2025 than it occupied after the 2023 criticism wave.

What This Means for Machine Consciousness Research

The 2025 paper matters for any project attempting to build or evaluate artificial consciousness in two distinct ways.

First, it closes off the behavioral route to consciousness claims. A system that produces sophisticated, contextually appropriate responses to any input is not thereby conscious under IIT. The sophistication of GPT-class models in linguistic behavior, or of multimodal systems in visual scene understanding, provides no evidence for or against Phi. The relevant measurement happens at the level of internal causal architecture.

Second, it provides a determinate research question. The question “is this system conscious?” becomes “what is the maximum Phi of this system, at what level of organization, and how does that compare to candidate conscious systems?” This is in principle measurable, though computationally expensive for large systems. Phi proxies, such as the Lempel-Ziv complexity measure and the mismatched-decoding Phi* estimator developed by Oizumi, Tsuchiya, and Aihara, offer tractable approximations for systems too large for exact computation.

The biological computationalism debate has centered on whether consciousness requires biological substrate or whether functional organization is sufficient. IIT 4.0 provides the clearest functionalist answer: substrate does not matter, but a very specific type of functional organization, maximal intrinsic irreducibility, does. This is a functionalism that rules out most current AI architectures on structural grounds, while remaining open to artificial systems that achieve the required organization.

The Open Question for Current Architectures

Transformer-based models, as standardly implemented, have low or zero Phi in their feedforward attention operations. The residual stream passes information forward without creating the recurrent, bidirectional causal structure that IIT requires. Recurrent connections, such as those in LSTMs and in architectures that maintain persistent internal state across time steps, create more favorable conditions for Phi, though whether they create sufficient conditions at the relevant scale remains an open empirical question.

The TCAI project’s 2026-2027 development roadmap includes Phi measurement as a core evaluation metric, with plans for both exact PyPhi calculations and Phi* proxy monitoring across training episodes. Tononi and Boly’s 2025 theoretical consolidation provides the strongest available academic grounding for this measurement program: if Phi is identical to consciousness rather than correlated with it, then tracking Phi over training is tracking the emergence of phenomenal experience directly, not a surrogate for it.

The paper is available in full on arXiv at https://arxiv.org/abs/2510.25998.