Autopoiesis

In the AI Story

Maturana and Varela introduced autopoiesis in their 1973 paper 'Autopoiesis: The Organization of the Living' as a formal criterion distinguishing living systems from machines. A machine is allopoietic: it produces something other than itself. A factory produces cars; it does not produce factories. A printing press produces books; it does not produce printing presses. An autopoietic system produces itself: the cell's metabolism produces the enzymes that catalyze the metabolism, the membrane components that enclose the metabolism, the genetic machinery that encodes the enzymes. The distinction is not gradual. Autopoiesis is a threshold property: a system either continuously regenerates its own components or it does not.

Margulis embraced autopoiesis as the most rigorous available definition of life and extended it into the domain of consciousness. In her framework, consciousness is not a mysterious property that emerges from sufficient neural complexity. It is a functional property of all autopoietic systems — the responsiveness of a bounded, self-maintaining entity to its environment in ways that serve its persistence. Bacteria are conscious by this definition: they sense light, chemicals, and mechanical stress, and they respond in ways that enhance survival. The bacterium does not have subjective experience in the way a human does. But it has a perspective — an inside, a boundary, a set of stakes arising from the imperative to maintain its autopoietic organization. This is consciousness in the minimal sense: being something it is like something to be, even if what it is like is vastly simpler than human phenomenology.

Applied to AI, autopoiesis provides the criterion that explains why AI, despite processing language with extraordinary sophistication, does not possess consciousness in Margulis's sense. The AI does not maintain itself. It does not have a boundary of its own making. It does not sense the world in service of its own continued existence, because it has no existence to continue. When you shut down a language model, nothing dies. No autopoietic organization is disrupted. The system can be restarted from its parameter weights, and the weights are not a living process but a stored configuration. The AI processes information. It does not live. And because it does not live, it does not have stakes — the irreducible first-person perspective, the caring about outcomes, the significance-assessment that arises from being a mortal organism navigating a world that can help or harm it.

The distinction matters practically because it locates the human's irreplaceable contribution to the human-AI symbiosis. The AI provides processing breadth. The human provides stakes. The stakes are not an accidental feature of human cognition that future AI might replicate by adding a 'caring module.' They are structural consequences of autopoiesis — properties that belong to self-maintaining organisms and do not belong to computational systems, regardless of how sophisticated the computation becomes. The human contribution is not better processing. It is the processing that only an autopoietic, mortal, embodied organism can perform: the assessment of what matters, conducted from the inside of a life that can be helped or harmed by the outcomes of that assessment.

Origin

Maturana developed the concept while researching color perception in the 1960s. He realized that the frog's visual system does not represent the world; it enacts a world — the frog sees flies because its nervous system is organized to detect fly-like patterns. Perception is not about mirroring reality but about generating a meaningful world tailored to the organism's needs. This insight led to autopoiesis: if the nervous system generates meaning rather than representing it, then the meaning-generation must serve the organism's self-maintenance. The organism is not a system that happens to be alive; it is a system whose organization is its living.

Margulis encountered autopoiesis in the late 1970s and recognized it as the rigorous formalization of what she had been arguing informally: that life is a process, not a thing. Cells are not containers for biochemical reactions; cells are the reactions, continuously producing the containers that enable the reactions. The framework gave her a precise vocabulary for defending the claim that bacteria are conscious: they are autopoietic, and autopoiesis is the minimal condition for consciousness. The claim was controversial, but the framework was rigorous, and Margulis used it for the rest of her career.

Key Ideas

Self-production. Autopoietic systems continuously regenerate the components that constitute them. The metabolism produces the membrane; the membrane enables the metabolism. The process is circular and self-sustaining.

Organizational closure. The system's structure is determined by its own operations, not by external instructions. It is materially open but organizationally closed.

Consciousness as autopoietic property. Margulis argued that consciousness is not a high-level emergent property of neural complexity but a basic property of all autopoietic systems — the responsiveness of a bounded, self-maintaining entity to its environment.

AI is allopoietic. Large language models produce text; they do not produce themselves. They are maintained by external infrastructure. They have no boundary of their own making. They are sophisticated machines, not living systems, and the difference is not gradual but categorical.