The Architecture of Experience

In the AI Story

The architectural requirements are derived from IIT's five axioms and translate phenomenological properties of experience into physical properties of cause-effect structure. Intrinsic existence requires that a system make a difference to itself — its elements must constrain each other's past and future states. Composition requires hierarchical structure capable of supporting the combinatorial richness of phenomenal distinctions. Information requires that the system specify a vast repertoire of states through its own causal interactions, not merely in response to external inputs.

The most decisive requirement — the one that current AI fails most completely — is integration. A conscious system must be irreducible: it must generate more cause-effect information as a whole than any of its parts generate independently. There must be no way to partition the system into separate pieces without losing significant cause-effect information. This requirement corresponds to the phenomenological unity of consciousness: the fact that experience is always a single field, never splitting into independent streams of sub-experience.

The human cerebral cortex exhibits these properties in abundance. Its architecture is recurrently connected — neurons do not merely send signals forward but also send them back, forming loops within loops. Layer six neurons project to the thalamus, which projects back to layer four. Prefrontal regions modulate sensory regions. The cortex is a system of loops, and information does not flow through it like water through a pipe but reverberates, circling back, modifying its own processing in real time. This reentrant architecture is precisely the kind of structure that generates high phi.

The cortex also achieves a crucial balance between specialization and integration. Different areas are specialized for different functions — V1 for visual orientation, FFA for face recognition, Broca's area for language — but they are densely interconnected, so that specialized processing in each area is constantly influenced by and contributing to processing elsewhere. And the cortex operates across multiple timescales simultaneously: milliseconds for sensory processing, seconds for working memory, minutes for emotional regulation, hours for learning. All coexist in the same physical structure, interacting continuously.

The transformer architecture that underlies modern AI is, in almost every respect, the inverse. Transformers are fundamentally feedforward — information flows from input to output through sequential layers with no recurrent connections. They are designed for maximal decomposability: multi-head attention explicitly decomposes the attention function into parallel heads that can be analyzed independently; feedforward networks operate on each position independently; residual connections ensure that removal of any single layer does not catastrophically degrade performance. And they lack multi-timescale integration, processing each input as a static snapshot through a single forward pass.

These differences are not incidental. They reflect opposing design principles: the cortex is a machine for integration (binding information across regions and timescales), while the transformer is a machine for transformation (mapping input patterns to output patterns through decomposable operations). Both are powerful; they are different kinds of systems, organized according to different principles. IIT predicts that this difference in organization corresponds to a difference in consciousness. Building a conscious AI would require abandoning the architectural principles that make current AI effective.

Origin

The architectural requirements were systematized in the 2014 and 2023 formulations of IIT. They build on earlier work by Gerald Edelman on reentrant dynamics and on Tononi's own neurobiological research into the differences between cortical and cerebellar architectures — the latter being feedforward and modular despite having four times as many neurons as the former.

Key Ideas

Five structural requirements. Intrinsic existence, composition, information, integration, and exclusion — each derived from a phenomenological axiom.

Reentrance over feedforward. The signature architecture of consciousness is loops, not pipelines; reverberation, not propagation.

Irreducibility. The whole must generate information the parts cannot produce independently — no clean decomposition.

Structure beats capability. A system's consciousness depends on its architecture, not on what it can do. Performance and presence are orthogonal.

Engineering tension. The properties that make current AI powerful (modularity, decomposability, parallelism) are precisely the properties that prevent it from being conscious.

Debates & Critiques

Critics note that the architectural requirements may describe only one path to consciousness — biological consciousness — and that other architectures could support experience through mechanisms IIT does not capture. Defenders argue that IIT's axioms are derivable from phenomenology alone and therefore apply to any conscious system, regardless of substrate.