An Algorithmic Information Theory of Consciousness

Consciousness is compression: this paper argues that what it feels like to be aware of something is what it looks like, from the inside, when a brain runs a good predictive model of its world.

The core bet is simple. A brain is a model-building machine. It takes in sensory streams, builds compact internal representations (models) that predict what comes next, and compares those predictions against reality. The richer and more compressive those models — the shorter the program that generates them, in the language of Kolmogorov complexity — the richer the conscious experience. A thermostat reacts; it doesn't model. A brain that builds hierarchical, self-referential models of its own sensory streams and actions is doing something qualitatively different, and that difference is what consciousness is.

The formal machinery comes from algorithmic information theory (AIT). Kolmogorov complexity K(x) is the length of the shortest computer program that produces a string x — a distribution-free, objective measure of how much structure is in a sequence. The paper grounds consciousness in this: an agent's experience is structured to the degree it has access to compressive, accurate, wide-coverage models of its input/output streams. Three dimensions follow naturally — model simplicity (how short the description), breadth (how much of the sensory stream it covers), and accuracy (how well predictions match reality). Richer consciousness means better scores on all three. The paper also identifies mutual algorithmic information between agent and world as a necessary, though not sufficient, condition: you need to be coupled to the world, but coupling alone doesn't make you conscious.

The architecture that does this work has three modules. A Modeling Engine builds and updates the internal model. An Objective Function evaluates how good the current model state is — think homeostasis, valence, reward. A Planning Engine selects actions to improve that score. The crucial functional locus is the Comparator: the moment the model's prediction is checked against incoming data. The paper's central claim is that structured conscious experience arises at the Comparator — it is the act of successfully matching model to world that constitutes experience, not the model sitting passively in memory.

The paper is honest that Kolmogorov complexity is uncomputable in general, so it proposes practical proxies: Lempel-Ziv-Welch compression ratios on EEG, the Block Decomposition Method on neural time series, binocular rivalry paradigms (the brain should "choose" the more compressible image as the dominant percept), and TMS perturbation studies to disrupt modeling networks and watch complexity metrics fall alongside reported consciousness. These are real, testable predictions, not just philosophical gestures. The source notes that later papers in the same research program have begun empirically validating these methods, though the mechanistic neural detail — which circuits implement the Comparator, how error feeds back — is left for subsequent work.