The Markov Blanket

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what it means to remain yourself through the encounter with AI—not as a sentimental concern about identity but as a practical question about agency, capability, and the development of judgment over time. The Markov blanket concept gives this question a formal structure. The self, on Friston’s account, is not a substance or a soul but a statistical boundary maintained over time: the set of states that separate what is internal from what is external. A self that maintains its Markov blanket is a self that persists as a distinct entity through changing circumstances. A self whose internal states come to be systematically determined by an external model—not mediated by the blanket but bypassed by it—is a self whose blanket is collapsing.

This is not a metaphor. It is a precise description of what happens when a practitioner outsources the cognitive work that previously constituted their internal states. The writer whose language sense was an internal state—part of their own generative model of their domain—and who now defers that judgment to a language model has moved something from inside the blanket to outside it. The blanket itself may remain intact in a biological sense; the question is whether the states that constitute the practitioner’s professional self still have the conditional independence from external inputs that would make them genuinely internal.

Friston’s framework thus converges with K. Anders Ericsson’s finding about the conditions for representational development, and with automation dependence research, from a completely different direction: they all describe, in different vocabularies, the same structural risk of AI-assisted work. Ericsson calls it the failure to build mental representations. Automation dependence researchers call it skill atrophy. Friston calls it blanket erosion. The cycle calls it becoming less worth amplifying.

Origin

The Markov blanket concept originated in the graphical models literature, where it refers to the set of nodes that separate a given node from the rest of a Bayesian network: the node’s parents, its children, and its children’s parents. Given the states of its Markov blanket, a node is conditionally independent of all other nodes in the network. This technical definition was imported into neuroscience and theoretical biology by Friston, who argued that it captures something fundamental about what it means to be a separate entity.

The philosophical extension—from a technical tool in graphical models to a definition of individuation—appeared in Friston’s 2013 paper “Life as We Know It” in the Journal of the Royal Society Interface, and was most fully developed in the 2019 preprint “A Free Energy Principle for a Particular Physics.” The central move is to argue that anything that can be said to exist as a distinct thing must, by definition, possess a Markov blanket: a boundary that keeps its internal states from being directly determined by everything in the environment at once. To have a boundary, in this sense, is to be an entity.

The concept has been received with a mixture of enthusiasm and skepticism. Enthusiasts point to its elegant unification of biological individuation, cognitive selfhood, and the formal tools of graphical models. Skeptics argue that the definition is too permissive—any cluster of correlated variables can be given a Markov blanket, so the concept does not distinguish genuine selves from statistical artifacts—and that the move from statistical structure to ontological selfhood requires more argument than Friston supplies.

Key Ideas

Statistical individuation. The Markov blanket provides a mathematical answer to the question “what makes an entity distinct from its environment?” The answer is: conditional independence. An entity’s internal states are conditionally independent of the external environment, given the states of the blanket. This is not a metaphysical claim about substance; it is a statistical claim about information flow.

The blanket as the locus of active inference. The Markov blanket defines the boundary across which active inference operates. The system’s actions change the external states that impinge on the blanket; the sensory data that arrives at the blanket updates the internal generative model. Without a blanket, there is no systematic distinction between what is inside and what is outside, and therefore no coherent sense in which the system is acting on a world rather than simply participating in a field of causation.

Nested blankets and social selves. Friston extends the concept hierarchically: cells have Markov blankets; organs are assemblies of cells whose blanket structures are nested; organisms are assemblies of organs; social groups are assemblies of organisms. Each level of the hierarchy has its own blanket, and each can be analyzed in terms of free energy minimization at that level. The social implications—including the question of whether AI systems and human organizations share nested Markov blankets in ways that are politically and ethically significant—are largely unexplored.

The question for AI systems. Do large language models have Markov blankets in the sense required for genuine agency? The technical answer depends on how the architecture is analyzed. Friston’s argument is that the blanket relevant for intelligence is not the physical boundary of the server but the statistical structure that separates a system’s internal generative model from its environment in a way that persists over time and is maintained through action. Current systems, trained on static data and deployed as stateless query-responders, do not maintain this structure in the relevant sense—and this, Friston argues, is the architectural gap that separates them from genuine agents.