The World of SPLectrum
Home > Positioning > Subjects > Complex Adaptive Systems > Contested receptions
Contested receptions
What is contested within and around the CAS tradition — definitional questions, internal differences, external critiques, and the boundary between rigorous application and pop-science metaphor. These disputes are part of what CAS is in the world, not problems to be tidied away.
Definition
What counts as “complex” or “adaptive” is contested, and the slipperiness is structural. The definition has to be broad enough to cover immune systems, economies, ecologies, and ant colonies — but narrow enough that not everything qualifies. The “everything is a CAS” risk is real: if the definition is too broad, the concept loses discriminatory power and becomes a label rather than a tool.
Contested cases sit at the boundaries. Does biological evolution count as a CAS? The population of organisms adapts, but individual organisms do not change their genotypes through interaction — the adaptive unit is the population, not the agent. Does the brain count? Neurons interact locally and produce global patterns, but the system is not a population of independent agents in the usual sense. Does the global economy count as a single CAS or as many coupled ones?
Mitchell (Complexity: A Guided Tour, 2009) addresses the definitional problem directly. Her working criteria: a system of many interacting components, without central control, whose aggregate behaviour is emergent, adaptive, and not deducible from the components alone. She treats these as sufficient conditions rather than a sharp definition — a diagnostic rather than a boundary. The complex/complicated distinction often invoked in CAS-adjacent discourse (a jet engine is complicated; a rainforest is complex) captures something real but can flatten into slogan.
Holland vs. Kauffman
Two emphases within the tradition, coexisting at SFI rather than resolving. Holland centres the adaptive agent with internal models — the agent that learns, adjusts its strategies, and generates anticipatory behaviour. The research program this produces is agent-based modeling: specify agents, run simulations, observe what the population generates. The explanatory mode is bottom-up and generative.
Kauffman centres the population dynamics — autocatalysis, self-organization, order arising from network topology rather than from agents’ individual strategies. The research program this produces is mathematical and statistical: Boolean networks, NK landscapes, the conditions under which order arises “for free.” The explanatory mode is structural: what does the space of possible configurations look like, and where does the system sit within it?
The disagreement plays out in practice. Holland’s camp asks: what do the agents do, and how does their adaptive behaviour produce the macro-pattern? Kauffman’s camp asks: what constraints does the network topology place on possible outcomes, regardless of what individual agents do? Both produce genuine results; neither subsumes the other. The coexistence is productive — but it means that “CAS explains X” can mean quite different things depending on which camp is speaking.
The metaphor critique
John Horgan’s The End of Science (1996) charges complexity science with producing “ironic science” — work that is “more akin to literary criticism or philosophy than to science.” The specific claim: CAS generates suggestive metaphors and narratives about emergence, self-organization, and adaptation, but does not produce testable predictions or quantitative laws. The Santa Fe Institute, in Horgan’s reading, is brilliant at generating frameworks and terrible at generating results that can be falsified.
The defences come from several directions. Mitchell (Complexity, chapters 17–19) argues that the charge conflates prediction with explanation and that generative modeling — showing how patterns can arise — is the appropriate epistemic stance for non-linear, path-dependent systems. Mark Bedau’s work on artificial life epistemology argues that agent-based models produce a distinctive form of understanding: they reveal sufficient conditions for phenomena that are opaque to analytical methods. Naomi Oreskes’ broader argument about models — that verification and validation are impossible in open systems, and that models should be evaluated as tools for exploration rather than mirrors of reality — provides an adjacent defence from philosophy of science.
The critique has not gone away. It resurfaces whenever CAS vocabulary is applied without the modeling discipline behind it — which is frequently.
Chaos confusion
The conflation happens primarily in popular treatments and media coverage. “Chaos theory” and “complexity science” are used interchangeably; the butterfly effect is treated as a CAS concept; fractals are presented as the visual signature of complex systems. The conflation is understandable — both deal with non-linear dynamics, both emerged in overlapping periods, both were popularised in the same wave of popular-science publishing in the late 1980s and 1990s.
The technical distinction is clear. Chaos describes deterministic systems with sensitive dependence on initial conditions — the system’s behaviour is fully determined by its rules, but tiny differences in starting conditions produce divergent trajectories. CAS describes populations of adaptive agents whose strategies change through interaction — the system’s behaviour depends not just on initial conditions but on agents’ ongoing responses to each other and to the environment. The overlap is in non-linearity; the distinction is in adaptation.
The “complexity” umbrella that includes both traditions doesn’t help. It groups CAS with chaos, fractal geometry, network science, and information theory under a single label, which obscures the specific commitments of each. CAS researchers generally resist the grouping but benefit from the public attention it generates.
Management appropriation
The application of CAS vocabulary to management and organisational theory is genuinely contested. Ralph Stacey developed the most sustained application: organisations as complex responsive processes, with management reconceived as participation in ongoing interaction rather than control from above. His work takes CAS seriously as theory, not as metaphor.
Dave Snowden’s Cynefin framework classifies situations into four domains: clear (cause-and-effect obvious, best practice applies), complicated (cause-and-effect discoverable through analysis, expert practice applies), complex (cause-and-effect coherent only in retrospect, emergent practice through safe-to-fail experiments), and chaotic (no discernible cause-and-effect, novel practice needed). The framework is widely used in management consulting and organisational design. It preserves structural content from CAS — the complex domain is genuinely different from the complicated domain, and the response strategies differ accordingly.
Other applications are thinner. “Our organisation is a complex adaptive system” often functions as a licence to avoid structural analysis rather than to do it differently. The boundary between rigorous application and pop-science metaphor is not settled, and CAS researchers remain divided on whether the management uptake has been net positive or net dilutive.
Predictability
CAS models are typically generative — they show how patterns can arise rather than predicting specific outcomes. Joshua Epstein’s formulation is the sharpest statement of the stance: “If you didn’t grow it, you didn’t explain it.” Generative sufficiency — showing that specified micro-conditions can produce the target macro-pattern — is treated as a form of explanation in its own right, distinct from prediction.
Cases where prediction succeeds: epidemic models can bound outbreak trajectories given assumptions about contact rates and intervention timing. Traffic models can predict congestion patterns from road topology and demand profiles. Cases where prediction fails: long-run economic forecasting, ecosystem dynamics over decades, social tipping points. The difference is not random — predictability correlates with the system’s sensitivity to agent-level heterogeneity and the rate at which the landscape is reshaped by agents’ actions.
Some critics treat the generative stance as a retreat from scientific standards. If the model is not predicting, what would count as failure? The response: generative models fail when they cannot reproduce the qualitative target pattern under any plausible parameter range, or when they produce the pattern only under narrow, implausible conditions. The validation problem — multiple micro-specifications producing the same macro-pattern — is real and connects to the methodologies discussion. The debate is unresolved and probably unresolvable at the general level; predictability depends on the specific system and the specific question asked of it.