Pedro Mediano and Integrated Information Decomposition

Pedro Mediano’s contribution is a decomposition rather than a new theory. Integrated information, as originally formulated, is a single number expressing how much a system’s whole exceeds its parts. It does not say what kind of excess it is. Integrated Information Decomposition, written PhiID, separates that quantity into parts that are redundant, meaning the same content sits in several places at once, and parts that are synergistic, meaning the content exists only when regions are taken together and in none of them alone. Once the two can be told apart, a formal defect in the original measure becomes repairable and the brain data starts saying something specific.

The Defect It Repairs

The original measure of integrated information can come out negative. That is not a rounding artefact, it is a conceptual problem, because a quantity meant to express how much a whole exceeds its parts should not be able to report that the whole is less than their sum.

Mediano, Fernando Rosas and colleagues traced this to the original formulation subtracting one quantity from another without distinguishing the kinds of information each contains. Separating redundancy from synergy first makes the subtraction well defined, and the revised measure they propose, written PhiR, is demonstrably non-negative.

That is a substantive repair to the most cited quantity in consciousness science, and it has attracted far less attention than the philosophical disputes about the same theory. The separate objection that phi is uncomputable and mis-ranks simple systems, set out in Scott Aaronson and the expander graph objection to integrated information theory, is untouched by this work. What it removes is one of the clean formal complaints.

What the Brain Data Showed

The framework was then tested on real recordings. Andrea Luppi, Mediano, Fernando Rosas and colleagues published A synergistic workspace for human consciousness revealed by Integrated Information Decomposition in eLife in 2024, volume 12, article RP88173, with Robin Carhart-Harris, David Menon and Emmanuel Stamatakis among the co-authors.

Applying the decomposition to human neuroimaging produced a division of labour that maps onto networks already known from other work. Regions acting as gateways, gathering synergistic information from diverse modules, correspond to the default mode network. Regions acting as broadcasters correspond to the executive control network. Together the authors call them a synergistic workspace.

The test was what happens when consciousness goes. Under anaesthesia and in brain injury, the workspace loses its ability to integrate information, and the gateway regions are affected most. That is a directional prediction, made on a quantity derived from theory rather than fitted to the result.

Why This Is a Bridge Between Rivals

The finding is more interesting than either camp usually produces, because an information theoretic method associated with integrated information theory went looking and found a global workspace.

Those two theories have been treated as rivals for two decades, and the adversarial collaboration built to separate them produced results neither side accepted as decisive, documented in the adversarial test of IIT against global neuronal workspace. This work suggests part of the disagreement was vocabulary. Synergistic information is what a workspace holds, so measuring integration properly reveals a broadcast structure rather than competing with one.

That matches the pattern Graziano argued for on entirely different grounds, that several major theories describe one mechanism from different angles, covered in Michael Graziano and the attention schema theory of AI consciousness. Where each theory sits is set out in the index of consciousness theories and what each predicts about AI.

What It Means for Artificial Systems

The decomposition is substrate neutral and computable on any system whose components can be recorded over time, which includes artificial networks. That makes it one of very few consciousness adjacent measures that can be run on a model rather than argued about.

The prediction it supports is specific. A system whose information is mostly redundant tolerates damage well and is not integrated in the relevant sense. A system whose information is mostly unique is a set of specialists with no shared state. What the theories ask for is synergy, content existing only in the joint state, and synergy is measurable.

Applied to a transformer, the question becomes whether attention produces genuine synergy across heads and layers or distributes redundant copies of the same content. That is an empirical question with a defined method, and better formulated than asking whether the architecture has a workspace. The same framing has been applied to multi-agent systems in swarm cognition and global workspace structure.

Why Synergy Is the Interesting Quantity

Of the three components, synergy is the one that carries the theoretical weight, and it is worth being precise about why.

Redundancy is what engineering usually optimises for. Storing the same content in several places means no single failure destroys it, which is why biological systems are full of redundancy and why it is a poor candidate for what consciousness consists of. A backup is not an integration.

Unique information is what specialisation produces. A module that alone carries some content is doing its job, and a system made entirely of such modules is a well designed pipeline with no shared perspective anywhere in it. That is close to a description of most software.

Synergy is the only one of the three that cannot be attributed to any part. The content exists in the joint state and vanishes when the components are considered separately, which is the property both integrated information theory and global workspace accounts have been reaching for with different vocabulary. Framing it this way also explains why the two theories kept producing compatible results while their proponents argued: they were both tracking synergy, one by measuring irreducibility and the other by looking for broadcast.

There is a caution attached. Synergy is a statistical property of a set of time series, and a system can have a great deal of it for reasons that have nothing to do with anything mind-like. Two coupled oscillators produce synergistic information. The measure narrows the question without answering it.

The Limits

Decomposing integrated information does not establish that integrated information is the right thing to measure. If synergy is not what consciousness consists of, measuring it better changes nothing, and every argument against the underlying theory applies with full force to the improved version.

The method is also computationally demanding, and the number of terms grows quickly with the number of components, so applications so far use coarse parcellations rather than fine grained recordings. That is a practical limit rather than a conceptual one, and it is the same constraint that has kept phi itself out of routine use.

Comparison to The Consciousness AI

This is the measure most directly applicable to the architecture described on the architecture page, and the reason is the structure of the Global Workspace layer.

That layer takes bids from specialist modules, ignites through a sigmoid non-linearity, and broadcasts back to all of them across five to ten reentrant cycles. Every module is instrumented and every state is recorded, which is the condition the decomposition requires and which biological recordings only approximate.

Running PhiID across those modules would say whether the broadcast produces synergistic information or redistributes redundant copies. That is a measurable property of the running system rather than an architectural claim, and it would be a real test rather than a demonstration. It is not implemented, and it is the most concrete open measurement this project could add.

What Follows

Mediano’s work is the least philosophical contribution in this area and possibly the most useful. It repairs a formal defect, produces a directional prediction about anaesthesia that held, and supplies a method that runs on artificial systems without modification.

What it does not do is settle anything about experience. Synergy is a property of information flow. Whether a system with a great deal of it feels like anything is the question that survives every improvement to the measurement, and this framework was never built to answer it.