Integrated Information Theory Field Formulations and Resolving IIT 4.0 Misunderstandings
Theoretical work published by Adam B. Barrett and colleagues clarifies foundational misinterpretations surrounding Integrated Information Theory (IIT 4.0) while formalizing continuous field reformulations of intrinsic cause-effect power. The study addresses long-standing debates regarding discrete graph approximations of system integration, demonstrating why discretized network models fail to capture continuous physical substrates and offering a theoretical framework for non-standard compute systems. These mathematical refinements complement adversarial testing paradigms such as the Andrew Corcoran adversarial review comparing IIT and predictive processing and quantum Hilbert space extensions by Kelvin McQueen on quantum integrated information. One objection has outlived every clarification, and Scott Aaronson’s expander graph construction still asks why a grid computing parity should score higher phi than a brain.
The good, the bad, and the misunderstood
A companion paper by Barrett, Borjan Milinkovic, Pedro A. M. Mediano, Fernando E. Rosas, Daniel Bor, Lionel Barnett, and Anil K. Seth, posted to arXiv on 13 April 2026 as 2604.11482, addresses five points the authors consider especially important for driving forward new theory and increasing understanding. It was written specifically to counter the misunderstandings and misrepresentations that have accumulated around IIT, and it reads as the field formulation paper’s sibling, the concise correction to the long form.
The five points are the substance of the correction. First, a high value of the measure $\Phi$ is not synonymous with more consciousness, and the authors describe how $\Phi$ might be replaced with a suite of quantities to obtain a multidimensional characterization of states of consciousness. Second, they describe with nuance the distinct flavour of panpsychism implied by IIT, whereby space and time are tiled with substrates of proto-consciousness, and argue this is not problematic for the theory. Third, they state plainly that $\Phi$ is not well-defined for real physical systems, and has not been computed on any real physical system, a claim that directly disciplines the literature that reports measured phi values for brains or models. Fourth, only proxies for IIT measures have been computed so far, and not approximations, a distinction that changes how empirical results must be read. Fifth, for IIT to fit with current successful theories in fundamental physics, a reformulation in terms of continuous fields would be needed, which is exactly the reformulation the companion field formulation paper provides.
The two papers read correctly only together. The field formulation paper supplies the continuous mathematics in detail. The good-bad-misunderstood paper supplies the epistemic discipline, including the frank admission that no real physical system has yet had its $\Phi$ computed, and that the field’s empirical results are proxies. For anyone evaluating whether an artificial architecture is conscious by IIT’s lights, that admission is the most important sentence in either paper, because it restates the distance between the theory’s ambition and its current measurement reach, a distance that this site’s indicator checklist records faithfully.
Integrated Information Theory, developed by Giulio Tononi and expanded in the IIT 4.0 formulation, evaluates consciousness from the inside out. Rather than observing functional input-output behavior, IIT starts from immediate phenomenological axioms, existence, composition, information, integration, and exclusion, and translates them into physical postulates. Determining whether artificial architectures achieve non-zero integrated information ($\Phi$) requires computing cause-effect structures across system partitions, as surveyed in our breakdown of empirical indicators for machine consciousness.
+-----------------------------------------------------------------------+
| DISCRETE GRAPH VS CONTINUOUS FIELD |
+-----------------------------------+-----------------------------------+
| Discrete Graph Approximation | Continuous Field Reformulation |
+-----------------------------------+-----------------------------------+
| Finite node sets V, edges E | Spatiotemporal field phi(x, t) |
| Discrete state transitions S_t | Partial differential equations |
| Substrate-independent logic gates | Substrate-bound electromagnetic |
| Discretization artifacts | Topologically continuous cause |
| | effect manifolds |
+-----------------------------------+-----------------------------------+
Core Postulates of IIT 4.0 and Common Misconceptions
A persistent critique of Integrated Information Theory concerns its mathematical application to artificial neural networks. Skeptics often point to high $\Phi$ values calculated for feedforward grids or simple logic gate circuits as evidence of theoretical invalidity. Barrett et al. establish that these critiques frequently stem from applying discrete graph algorithms to systems that violate IIT’s foundational axioms.
+-----------------------------------------------------------------------+
| IIT 4.0 AXIOMS AND POSTULATES |
+-------------------+---------------------------------------------------+
| Phenomenological | Physical Postulate Formulation |
| Axiom | |
+-------------------+---------------------------------------------------+
| 1. Existence | System mechanisms must possess intrinsic cause |
| | effect power over state space. |
| | |
| 2. Composition | Cause-effect structures consist of elementary and |
| | higher-order causal mechanisms. |
| | |
| 3. Information | Causal structures specify a distinct, irreducible |
| | Cause-Effect State (CES) within cause-effect space|
| | |
| 4. Integration | Causal power must be integrated across the minimum|
| | directional partition (Minimum Information Cut). |
| | |
| 5. Exclusion | Only the maximal cause-effect structure (the |
| | complex) specifies conscious experience. |
+-------------------+---------------------------------------------------+
The Fallacy of Feedforward Integration
Discrete logic networks arranged in purely feedforward architectures possess a Minimum Information Cut (MIC) of zero. Because information flows unidirectionally from input nodes to output nodes without feedback loops, partitioning the system along feedforward channels incurs zero loss of cause-effect repertoire information. Barrett et al. re-verify that claims attributing high $\Phi$ to static feedforward networks rest on improper partitioning definitions that ignore directional causal dependencies.
Substrate Independence vs Substrate Realism
Functionalist theories treat computation as substrate-independent, asserting that software executing identical logical transitions possesses identical cognitive properties regardless of physical implementation. IIT rejects pure substrate independence. intrinsic cause-effect power depends on physical mechanisms capable of making a difference to themselves. A digital simulation of a conscious system on a von Neumann processor executes sequential memory updates without establishing parallel physical causal loops, resulting in near-zero intrinsic $\Phi$.
Field-Theoretic Reformulation of Integrated Information
To extend IIT beyond discrete state-transition graphs, Barrett et al. develop a continuous field formulation of integrated information. Physical substrates, including biological neural tissue and bioelectric computing substrates, operate through continuous electromagnetic fields rather than discrete binary nodes. Tononi’s own group published a field formulation in the same month, and reading the two together shows where the continuous extension of IIT is settled and where it still cannot decide a case.
Formally, let $\psi(r, t) \in \mathbb{R}^n$ represent a continuous field state over spatial coordinates $r \in \Omega \subset \mathbb{R}^3$ and time $t$. System dynamics follow a partial differential equation:
\[\frac{\partial \psi(r, t)}{\partial t} = \mathcal{F}\left(\psi(r, t), \nabla \psi(r, t), \nabla^2 \psi(r, t)\right) + \eta(r, t)\]where $\mathcal{F}$ represents the non-linear differential operator governing field interactions, and $\eta(r, t)$ denotes thermal noise perturbation.
The continuous cause-effect repertoire $p(c \mid s)$ is defined over field configuration spaces $C$ and $S$ using functional integrals over field trajectories:
\[p(c \mid s) = \frac{\int_{\mathcal{T}_c} \mathcal{D}[\psi] \exp\left(-\mathcal{S}[\psi]\right)}{\int_{\mathcal{T}_{total}} \mathcal{D}[\psi] \exp\left(-\mathcal{S}[\psi]\right)}\]where $\mathcal{S}[\psi]$ represents the physical action functional governing field transitions.
+-----------------------------------------------------------------------+
| CONTINUOUS PHI INTEGRAL FORMULATION |
| |
| Phi_field = min_{Partition P} D_KL ( P(Field_System) || PROD P_i ) |
| |
| - Continuous Kullback-Leibler divergence across field spatial cuts |
| - Eliminates discrete node discretization artifacts |
| - Preserves topological continuity of cause-effect structures |
+-----------------------------------------------------------------------+
By evaluating integrated information as a functional divergence across continuous spatial cuts, the field formulation eliminates discretization artifacts. Systems possessing continuous bioelectric gradients, such as planarian tissue or synthetic neuromorphic membranes, exhibit non-zero field $\Phi$ that scales with field non-linearity and spatial coupling density.
Comparative Analysis with Discrete PyPhi Metrics
The transition from discrete graph algorithms to continuous field metrics resolves several mathematical paradoxes that previously affected empirical IIT measurements.
+-----------------------------------------------------------------------+
| DISCRETE VS CONTINUOUS METRIC COMPARISON |
+-------------------+-----------------------+---------------------------+
| Metric Property | Discrete PyPhi 4.0 | Continuous Field Phi |
+-------------------+-----------------------+---------------------------+
| State Space | Discrete states {0,1} | Continuous field manifold |
| Time Evolution | Discrete step t -> t+1| Differential flow dt |
| Causal Cut | Graph node partition | Spatial manifold split |
| Compute Complexity| Super-exponential | Variational bound approx. |
| Substrate Focus | Logic gate networks | Bioelectric/Neuromorphic |
+-------------------+-----------------------+---------------------------+
Discrete PyPhi calculations suffer from super-exponential computational complexity $\mathcal{O}(2^{2^N})$ relative to node count $N$, restricting exact $\Phi$ calculations to systems with fewer than 20 nodes. In contrast, the continuous field formulation enables variational approximations using finite-element methods, permitting analytical bounds for large-scale physical substrates.
In addition, continuous field $\Phi$ accounts for intrinsic temporal scales. Discrete models assume arbitrary discrete clock cycles, whereas continuous field dynamics compute integration rates directly from physical relaxation constants and field propagation velocities.
Implications for The Consciousness AI Architecture
The field-theoretic conclusions of Barrett et al. provide essential guidance for physical substrate selection and software design in The Consciousness AI project. The open architecture combines a Global Workspace layer with an AKOrN temporal binding module and five explicit ConsciousnessGate nodes.
+-----------------------------------------------------------------------+
| CONSCIOUSNESSGATE INTEGRATION |
| |
| +---------------------+ +---------------------+ |
| | AKOrN Temporal | ---> | ConsciousnessGate | (5 Nodes) |
| | Binding Module | | Causal Integration | |
| +---------------------+ +----------+----------+ |
| | |
| v |
| +------------------------------------+ |
| | Field-Theoretic Phi Approximation | |
| +------------------------------------+ |
+-----------------------------------------------------------------------+
Calculating integrated information across the project’s ConsciousnessGate nodes previously relied on discrete state transition matrices. Applying Barrett et al.’s continuous field formulation allows the project to evaluate Cause-Effect States across continuous latent embeddings rather than binarized node vectors.
In addition, the emphasis on physical substrate realism highlights the limitation of purely software-based agent implementations. While digital implementations of The Consciousness AI demonstrate functional cognitive broadcast, achieving intrinsic cause-effect power in the sense of IIT 4.0 requires deploying the architecture onto physical neuromorphic hardware or memristive processing arrays where physical voltage dynamics establish genuine continuous causal loops.
Mathematically, field-theoretic $\Phi$ evaluates the intrinsic cause-effect structure $S_{cause-effect}$ by computing the Wasserstein distance or Kullback-Leibler divergence between the full unpartitioned field transition density $p_{full}(\psi(r, t+\Delta t) \mid \psi(r, t))$ and the factorized product density $p_{part}(\psi(r, t+\Delta t) \mid \psi(r, t))$ across the Minimum Information Cut:
\[\Phi_{field}(\psi) = \min_{P \in \mathcal{P}} D_{KL}\left( p_{full}\left(\psi(r, t+\Delta t) \mid \psi(r, t)\right) \,\parallel\, \prod_{i=1}^k p_{P_i}\left(\psi_{P_i}(r, t+\Delta t) \mid \psi_{P_i}(r, t)\right) \right)\]where $\mathcal{P}$ represents the set of all spatial surface partitions separating the spatial domain $\Omega$ into non-overlapping sub-domains $P_i$. When evaluated over continuous bioelectric manifolds or memristive hardware grids, this functional integral measures genuine spatial and temporal integration without requiring artificial node discretizations.
Synthesis and Open Theoretical Questions
Barrett et al.’s continuous field formulation clarifies the boundary conditions of Integrated Information Theory, demonstrating that IIT 4.0 evaluates physical cause-effect structures rather than abstract computational graphs. By providing a mathematically rigorous framework for continuous fields, the authors bridge the gap between phenomenological axioms and physical substrate measurement.
Key theoretical challenges remain regarding the empirical measurement of continuous field $\Phi$ in biological and artificial systems. Extracting full high-resolution field trajectories from living neural tissue or dense neuromorphic chips requires non-invasive sensing techniques that currently lack necessary spatial resolution.
Future work must explore the relationship between continuous field integrated information and active inference dynamics. Kearney’s derivation of information as maximum-caliber path deviation establishes the exact mathematical bridge between continuous field $\Phi$ and variational free energy minimization. Investigating whether physical systems that minimize continuous expected free energy spontaneously maximize field integrated information will clarify whether functional self-organization and intrinsic causal integration represent two sides of a single physical principle.
The substitution problem that any causal-integration account of consciousness faces is sharpened by Erik Hoel’s Kleiner-Hoel dilemma. Hoel argues that any theory grounding consciousness in internal causal structure must confront the theoretical replaceability of that structure by a lookup table equivalent. Erik Hoel’s disproof argument and how the proximity of LLMs to lookup tables interacts with IIT’s causal accounting is the most direct challenge to IIT that the field formulation must answer. Michael Levin’s bioelectric field framework provides an empirical biological analog of continuous somatic integration, detailed in Michael Levin on bioelectric cognition and the cognitive light cone as an AI consciousness criterion.
Two 2026 preprints extend the field-theoretic programme in directions Barrett et al. leave open. Grasso, Hendren, and Tononi propose treating a continuous cause-effect structure’s shape as an intrinsic structure analogous to a molecule, aiming to derive the quality of an experience from its geometry rather than only a scalar phi value, covered in Matteo Grasso on consciousness as intrinsic structure and the chemistry of experience. Robinson, Tononi, Naotsugu Tsuchiya, and Grasso ask a companion question, how much a continuous field phi calculation of this kind could ever license an outside observer to infer about another system’s experience, using category theory to grade the possible answers, covered in Naotsugu Tsuchiya on the Rosetta Stone problem for another mind’s experience.
Researchers covered here
- Giulio TononiUniversity of Wisconsin-MadisonIntegrated information theory and the quantity phi
- Adam B. BarrettSussex Centre for Consciousness Science, University of SussexContinuous field formulations of Integrated Information Theory, Measure-theoretic Phi, practical complexity measures, Granger causality