CONCEPT

Open System

Bertalanffy’s foundational distinction between closed systems that wind down to equilibrium and open systems that maintain themselves far from equilibrium through continuous exchange with their environment—the lens that dissolves the fiction of a deployed AI model as a finished, certifiable artifact.

An open system is Ludwig von Bertalanffy’s most productive concept, and the one that most directly challenges how we think about deployed AI. A closed system is governed by the second law in its simplest form: left alone, it runs down, entropy increases, disorder grows, and the system approaches equilibrium, which for a living thing means death. A living organism avoids this fate not by violating the law but by being a different kind of object with respect to it. “Every living organism,” Bertalanffy wrote, “is essentially an open system. It maintains itself in a continuous inflow and outflow, a building up and breaking down of components.” The organism holds itself in a steady state far from equilibrium through the sustained exchange of matter and energy with its environment; cut off the inflow and the organism slides toward the equilibrium it had been avoiding, which is death. The concept’s application to AI is structural rather than metaphorical. A deployed large language model is habitually described as a fixed artifact—a file of frozen weights, a finished product—and in the narrow sense of parameter constancy this is true. But a model’s behavior is constituted in exchange: it does nothing until something flows through it, and the behavioral steady state it maintains depends on the continuing flow of prompts, contexts, and environments that are themselves changing. When the environment drifts—when users evolve, the world’s text changes, the deployment context shifts—the system’s effective behavior drifts with it, even though no weight has been touched. The closed-system intuition that a trained model’s behavior is certified at deployment is the intuition Bertalanffy spent his career dismantling, and it continues to mislead AI safety thinking.

In the [YOU] on AI Field Guide

The cycle’s treatment of AI alignment and safety is directly sharpened by this concept. The dream that a model’s values and tendencies could be fixed at training and thereafter relied upon is, in Bertalanffy’s terms, the closed-machine dream: the hope for a thing that, once correctly built, stays correct. His biology says that no open system works this way. Homeostasis is not a setting dialed in at birth; it is an active, ongoing accomplishment, continually re-achieved against continual perturbation, and it can fail. The same structural point applies to alignment in deployed AI systems: it is not a property installed at training but a steady state that must be actively maintained against a shifting world.

The open-system lens also reframes the boundary of what counts as “the system.” A frontier model coupled to live web access, external tool calls, memory systems, and routing logic that processes its own intermediate outputs is constituted by its inflows and outflows in exactly the way Bertalanffy described for organisms. Its competence is not stored statically inside the weights; it is maintained dynamically through continuous exchange with an informational environment that includes the internet, a corpus of documents, a stream of user interactions, and a battery of tools. Bertalanffy would have recognized this configuration instantly and said: the interesting object is not the weights. It is the behavioral steady state the whole coupled system maintains.

The concept also illuminates what Lucy Suchman calls the open world: the unreduced, unrepresented reality within which human practitioners act and AI systems operate only on representations of. Bertalanffy’s open system is open precisely because it is in genuine exchange with the open world; a model, however capable, operates on descriptions of situations rather than situations themselves. The two frameworks converge on the same diagnostic: the system’s behavior is always a function of the exchange, and the exchange is always with something richer than any representation of it.

Origin

Bertalanffy developed the open-system concept in the 1920s and 1930s as a direct challenge to the application of equilibrium thermodynamics to living things. Classical thermodynamics, built for closed systems approaching equilibrium, could not account for the most basic features of biological organization: growth, self-repair, reproduction, the maintenance of internal complexity against an environment that would dissolve it. His 1932 Theoretische Biologie introduced the formal treatment of open systems as a thermodynamic class: systems that exchange matter and energy with their environment and can therefore maintain steady states far from equilibrium, sustaining internal organization through the continuous throughput of free energy.

The open system was explicitly distinguished from cybernetic feedback systems, which Bertalanffy regarded as a special case: a thermostat is a closed system that uses feedback to maintain a target state, but it does not grow, develop, or maintain itself against entropic decay. An organism is an open system that maintains far-from-equilibrium organization through metabolic exchange, not merely through feedback regulation. The distinction is important for AI: the training loop is pure cybernetics (objective, feedback signal, correction), but the capability that emerges from training is the Bertalanffian part—the self-organized internal structure that no one specified and that the feedback signal never named.

Key Ideas

Steady state vs. equilibrium. An open system maintains a dynamic steady state far from equilibrium through continuous inflow and outflow; a closed system approaches the equilibrium of maximum entropy. A deployed model in production is in a behavioral steady state that depends on the continuing exchange with its environment. Cut off the exchange (take the model offline, stop the prompts) and the behavior disappears, exactly as an organism deprived of metabolic exchange slides toward death. The behavior was never inside the weights; it was always in the relation between the weights and the environment.

Alignment as ongoing achievement. If a model’s behavior is a steady state maintained through exchange, then alignment cannot be a property certified once and trusted forever. The environment changes; the steady state the system maintains changes with it. A model aligned to 2024’s deployment environment may produce subtly different behavior in 2026’s environment, not because any parameter changed but because the environmental inputs through which the behavior is constituted have shifted. This is why AI safety requires monitoring and active maintenance rather than one-time certification.

The model is the whole coupled system. For the most capable deployed AI systems, the weights are only one component. The behavioral object is the model plus its tools, memory, retrieval systems, and real-time data access. Bertalanffy’s framework says: define the system boundary at the exchange interface, not at the artifact boundary. The meaningful system is the whole configuration that maintains the behavioral steady state, including all the environmental inputs that constitute it.

Debates & Critiques

The central debate the open-system concept provokes in AI concerns the practical implications for governance and safety. If alignment is an ongoing achievement rather than a certified property, what does this mean for deployment decisions, for liability frameworks, and for the institutions responsible for monitoring deployed systems? The closed-system intuition is deeply embedded in how software is regulated and how contracts for software products are structured; the open-system reality of deployed AI requires new legal and institutional frameworks that most jurisdictions do not yet have. A second debate concerns the degree to which the open-system analogy is literal versus structural. Bertalanffy himself was careful to note that a language model is not metabolizing, not self-repairing, not alive; the flows through it are informational rather than material-energetic, and granting it the full status of an organism would be a romantic error. The question is whether the structural analogy—behavior as exchange rather than property, steady state rather than certified output, environment-dependence rather than artifact-independence—is useful enough to justify the terminological investment. The evidence from deployed AI’s documented behavioral drift strongly supports the open-system frame over the closed-machine frame, regardless of whether the analogy is literal. Norbert Wiener’s cybernetics offers a complementary but narrower frame: the feedback loop explains how the training is governed but not how the resulting system maintains its steady state in deployment.

In the [YOU] on AI Field Guide

Origin

Key Ideas

Debates & Critiques

Related Entries

Further Reading