The concept does more conceptual work in the AI transition than any other framework from complexity science, because it locates AI within the continuity of natural adaptive systems rather than treating it as a discontinuous novelty requiring entirely new categories. [YOU] on AI returns repeatedly to the puzzle of AI’s simultaneous extraordinary capability and baffling fragility—the same system that produces a correct mathematical proof will fabricate a citation with perfect confidence. The CAS framework explains both features structurally: the capability reflects genuine schema compression of deep regularities in language and reasoning; the fragility reflects the shadows cast by a compression that operates on text-as-abstraction rather than on embodied encounter with a world where errors have consequences.
The framework also clarifies the human role in AI collaboration. Every human expert who evaluates AI output is functioning as what Gell-Mann would call a meta-schema: a compressed representation of years of feedback-driven domain experience that can detect when the AI’s schema has captured genuine deep structure and when it has retrieved a surface association. This meta-schema cannot be delegated to the AI, because the AI lacks the feedback mechanism that would allow it to evaluate the quality of its own compressions. The person who can tell when the output “sounds better than it thinks” is exercising the judgment that defines the tacit knowledge dimension of expertise—and that judgment is grounded in the CAS framework’s insight that not all regularities are equally deep.
Gell-Mann introduced the framework in its full form in his 1994 book The Quark and the Jaguar and in a series of papers written with colleagues at the Santa Fe Institute. The core intuition had been building since the Institute’s founding in 1984, when Gell-Mann assembled researchers from physics, biology, economics, and computer science around the hypothesis that complexity had universal principles waiting to be discovered. The explicit list of CAS examples—including “a computer learning to play chess” alongside biological evolution and cultural development—appeared in his 1994 Santa Fe Institute proceedings paper, establishing from the outset that artificial learning systems were within scope.
The concept built on earlier work by Norbert Wiener on cybernetics and feedback systems, but went further in specifying the role of schema compression—the idea that the adaptive system does not merely respond to feedback but builds an internal model that generalizes. This generalization capacity is what separates genuine adaptive systems from lookup tables, and it is the feature whose presence or absence in AI systems remains the central contested question of the current moment.
Schema Compression vs. Lookup Tables. Gell-Mann drew a sharp distinction between a system that genuinely compresses the regularities of its environment into a generative model and one that merely memorizes specific input-output associations. A genuine schema produces correct predictions about situations it has never encountered; a lookup table retrieves associations it has stored. The distinction is empirically real in current AI: when a model proves a theorem it was not trained on, compression is operating; when it fabricates a citation with perfect syntactic form, the lookup table is operating. The mixture varies by domain and task, which is why the question “Is the AI reliable?” always requires specifying “For what?”
The Embodiment Shadow and the Stakes Shadow. Gell-Mann identified two structural features of biological CAS that AI CAS lack. First, embodiment: a biological organism acquires information through a body that navigates physical consequences, making errors costly in ways that shape the depth of the schema. An AI acquires information through text—an abstraction that may or may not track the underlying reality it represents. Second, stakes: a biological system updates its schema because failure threatens its survival. An AI updates during training against a loss function defined by its creators—a proxy for reality that introduces a gap between what the schema is optimized for and what genuine reliability would require. These shadows are structural, not fixable by scale alone.
Coevolution of Schemas. In an ecosystem of complex adaptive systems, each system’s schema coevolves with the schemas of the other systems it encounters. As AI improves, the surface-regularity outputs become increasingly difficult to distinguish from deep-regularity outputs—the shadows become harder to see, not because they have disappeared but because the surface has become more polished. This coevolution places a continuous premium on the human meta-schema’s depth. The most dangerous regime is not when AI is obviously wrong but when it is almost right—when the schema captures enough deep regularity to be genuinely useful and enough surface regularity to be genuinely misleading, at a boundary invisible to anyone without deep domain expertise.