Society as a Grid: Emergence in Human Systems
In 1971, Thomas Schelling published a paper in the Journal of Mathematical Sociology (volume 1, pages 143–186) with the deliberately clinical title “Dynamic Models of Segregation.” He was an economist at Harvard, working in game theory and political economy. He had no connection to cellular automata research. He was not thinking about Conway’s Life, which had appeared in Scientific American the previous October.
And yet the paper he published was, in the most precise technical sense, a cellular automaton model.
Schelling’s model placed agents of two types — say, two racial groups — on a grid. Each agent had a preference: it would stay in its current location if some minimum fraction of its neighbors were its own type, and it would move to a more satisfactory location otherwise. The preference threshold was mild: Schelling typically used one-third. Each agent was content in a neighborhood where it was actually in the minority. The preference was not for dominance or exclusion — it was merely for not being entirely isolated.
When Schelling ran the model — first on a checkerboard, manually, pushing nickels and dimes around — the result was startling. Extreme, large-scale segregation emerged. The mild individual preference, operating through local interaction, produced a collective outcome that no individual intended, that exceeded what any individual preferred, and that was effectively irreversible once established. Neighborhood-level segregation was the emergent property of a system whose individual actors, by any reasonable standard, were not trying to create it.
This is the central phenomenon of the sociology cluster: social structure as emergent property of local rules.
Why This Is Not Just Analogy
When social scientists draw on CA models, they sometimes describe the relationship as analogical: society is like a cellular automaton, in the way a flock of birds is like a wave. The analogy is instructive but ultimately just a metaphor.
This description is too weak. Schelling’s model, Axelrod’s cultural dynamics model, the Clarke-Gaydos urban growth model, the SIR epidemic model in its spatial form — these are not analogies. They are cellular automata in the strict mathematical sense: arrays of agents, each holding a discrete state, each updating that state based on the states of neighboring agents according to a rule, with the global pattern being the aggregate of all local updates. The mathematical framework is the same. The questions asked are the same. The analytical tools — linear stability analysis, phase transitions, pattern formation — are the same.
The import of this is not that social systems resemble Conway’s Life. It is that the dynamics discovered in the abstract study of CA — emergence, phase transitions, sensitivity to initial conditions, the gap between local rules and global outcomes — apply directly, without translation, to social systems modeled this way. When a sociologist using a CA model finds that a small change in preference threshold produces a qualitative shift in segregation outcome, they are observing the same kind of phase transition that physicists observe in the Ising model or that Life researchers observe when they nudge rule parameters. The phenomenon is genuinely the same phenomenon.
Schelling’s Segregation Model: The Founding Document
Schelling’s 1971 paper is worth examining in detail, because its argument is more subtle than the simplified version that has circulated widely.
The paper considered two types of neighborhoods: those on a one-dimensional “line” (a long street, conceived as a string of positions) and those on a two-dimensional grid. In both cases, agents of two types were distributed with varying initial configurations, and agents who were below their preference threshold would move to a location where they were satisfied.
The key result was not merely that segregation emerged — it was the relationship between the preference threshold and the degree of segregation. At very low thresholds (agents satisfied even in heavily minority neighborhoods), the system remained integrated. But at surprisingly modest thresholds — often well below majority preference — the system collapsed to near-complete segregation. The transition was abrupt: small increases in the preference threshold produced disproportionately large increases in segregation.
This is a phase transition in the statistical mechanics sense: a sharp qualitative change in system behavior as a parameter crosses a critical value. Schelling observed it on a checkerboard with coins. Physicists observe the same phenomenon in the Ising model of ferromagnetism, where the analogue of the preference threshold is temperature and the analogue of segregation is magnetization. The mathematics is identical.
Schelling was honest about the limits of his model. He knew it abstracted away housing markets, discrimination (which is not the same as individual preference), historical path dependence, and economic constraints on mobility. He was not claiming to explain American residential segregation in its full historical complexity. He was identifying one mechanism — the amplification of mild individual preferences by local interaction — that contributes to the outcome. His contribution was to show that this mechanism alone was sufficient to produce substantial segregation, even without explicit discrimination.
The corollary — which Schelling emphasized and which is often forgotten — is that eliminating explicit discrimination does not eliminate this mechanism. A society without discriminatory laws but with mild individual preferences operating through local interaction can produce nearly the same outcome as a society with explicit discrimination. This is not a counsel of despair; it is a diagnosis that points toward specific interventions (changing the structure of the interaction, not just the law).
The Diffusion of Innovation: Rogers’ S-Curve as a CA
Everett Rogers’ 1962 book Diffusion of Innovations described how new ideas, technologies, and practices spread through populations. The model is famous for its S-curve: adoption starts slowly (early adopters), accelerates through the majority, and slows as laggards are reached. Rogers identified five adopter categories — innovators, early adopters, early majority, late majority, laggards — and described adoption as a social contagion process mediated by communication and social network effects.
This is structurally a cellular automaton. Individuals occupy nodes in a social network (the “grid”). Each individual has a state: not yet adopted, adopted. The transition from non-adoption to adoption is triggered by local social influence: an individual’s probability of adopting increases with the fraction of their social network that has already adopted. The S-curve is the aggregate output — the time series of total adoption — produced by this local rule applied simultaneously across the network.
The CA formulation of Rogers’ model makes several things clearer than the original description. The shape of the S-curve depends sensitively on the topology of the network: adoption spreads faster through networks with high clustering and slow through networks with sparse connections. The existence of the S-curve’s inflection point — the moment when adoption accelerates from slow to fast — corresponds to the CA analogue of a percolation threshold: the point at which the “infected” (adopting) region spans the network. These observations follow directly from CA theory and would be much harder to derive from Rogers’ verbal description alone.
Urban Growth: Cities as Cellular Automata
The application of CA to urban growth modeling is perhaps the most visually compelling in the social sciences. Satellite time-lapse imagery of city growth looks, strikingly, like the growth of patterns in Life: spread along corridors, sporadic emergence of outlying clusters, gradual filling-in of space between clusters, fractal texture at the boundary.
Roger White and Guy Engelen formalized this intuition in their 1993 paper “Cellular Automata and Fractal Urban Form: A Cellular Modelling Approach to the Evolution of Urban Land-Use Patterns” in Environment and Planning A (volume 25, pages 1175–1199). Their model treated each cell in a geographic grid as a land-use parcel that could transition between types — open space, residential, commercial, industrial — according to rules based on the land-use composition of neighboring cells. The model produced fractal land-use boundaries, hierarchical structure in the urban fabric, and dynamics that matched empirically observed patterns of urban growth.
Michael Batty and Yichun Xie extended the approach in their 1994 paper “From Cells to Cities” in Environment and Planning B (volume 21, pages 31–48), applying a CA model to the development of Savannah, Georgia, and showing that the historical cell-based structure of the city could be reproduced by local CA rules. Savannah’s famous grid of squares, each surrounded by residential and commercial blocks, is a literal cellular automaton in physical space — and the CA model predicted its development dynamics accurately.
Keith Clarke and Leonard Gaydos produced the most ambitious and widely used urban CA model. Their 1998 paper “Loose-Coupling a Cellular Automaton Model and GIS: Long-Term Urban Growth Prediction for San Francisco and Washington/Baltimore” in the International Journal of Geographical Information Science (volume 12, pages 699–714) described what became known as the SLEUTH model — named for its input layers: Slope, Land Use, Exclusion, Urban Extent, Transportation, and Hillshade. SLEUTH included four types of growth: spontaneous (isolated new development), diffusive (spread from existing urban areas), organic (filling in of existing urban areas), and road-influenced (growth along transportation corridors). After calibration against historical satellite imagery, the model produced quantitatively accurate projections of San Francisco Bay Area growth. Full details on urban CA models →
Cultural Evolution: Axelrod’s Model
Political scientist Robert Axelrod asked a question that sounds almost childish until you think about it carefully: why, in a world where people tend to become more similar to those they interact with, does cultural diversity persist? If social interaction causes convergence, the world should be culturally homogeneous. It is not.
Axelrod’s 1997 paper “The Dissemination of Culture: A Model with Local Convergence and Global Polarization” in the Journal of Conflict Resolution (volume 41, pages 203–226) addressed this with a CA model. Each agent on a grid had a cultural profile — a vector of features (like language, religion, dietary practice), each taking one of several possible values. Agents were more likely to interact with neighbors who already shared more features — homophily, the tendency to associate with similar others. When two agents interacted, they adopted a shared value for one feature, increasing their similarity. Convergence was local and continuous.
The surprising result: under many parameter settings, the system did not converge to a single global culture. It froze into a patchwork of distinct cultural regions, each internally homogeneous, between which there was too little similarity to initiate the interaction that would cause further convergence. The diversity was not maintained by any active force of differentiation — it was a frozen residue of the initial conditions, preserved by the local character of interaction. Cultures that were geographically isolated developed differently; geographic proximity did not guarantee eventual convergence once initial differences became large enough.
Axelrod’s model captures a phenomenon that purely mean-field models of cultural diffusion miss: the stabilization of diversity by geographic structure. When interaction is local, initial variation can become entrenched. The pattern of cultural diversity in the world — the extraordinary fragmentation of language and custom at fine geographic scales, alongside the broad homogenizing forces of trade and media — reflects exactly this tension between local convergence and global freezing that the CA model captures.
Epidemiology: SIR Models on the Grid
The SIR model — named for its three compartments: Susceptible, Infected, and Recovered (or Removed) — is the foundational model of infectious disease epidemiology. In its original mean-field form, developed by Kermack and McKendrick in 1927, it describes the evolution of the fraction of the population in each compartment as a function of the infection rate and recovery rate. It predicts epidemic curves — the characteristic bell-shaped time series of new infections — and identifies the basic reproduction number R₀ as the threshold parameter determining whether an epidemic occurs.
The spatial CA formulation of the SIR model replaces the mean-field assumption with explicit geography. Each cell on a grid is an individual (or a local population), in one of three states: S, I, or R. The update rule: susceptible individuals adjacent to infected individuals transition to infected with some probability; infected individuals recover with some probability. This is, structurally, Conway’s Life with three states instead of two, and a stochastic transition rule instead of a deterministic one.
The spatial SIR model produces qualitatively different results from its mean-field counterpart. Epidemic waves — traveling fronts of infection — emerge naturally in the CA formulation and are absent from the mean-field equations. Superspreader dynamics, where a single infected individual in a highly connected position causes disproportionate spread, appears when the grid topology reflects real social network structure. Local extinction is possible even when the global mean-field model predicts persistence.
These differences are not merely theoretical. The spatial heterogeneity of COVID-19 spread — the spectacular variation in outbreak timing, severity, and wave structure across geographic areas — was qualitatively better explained by spatial models than by mean-field models. The heterogeneous spatial spread of infection, the role of transportation corridors in connecting local outbreaks, the persistence of hot spots surrounded by cooler areas: these are CA phenomena.
The Deeper Lesson: Intentions and Outcomes Are Different Things
The sociology cluster has a unifying theme that is worth making explicit, because it is one of the genuinely important insights of 20th-century social science.
Complex social structures — segregated neighborhoods, culturally homogeneous regions, pandemic waves, sprawling city morphologies — are typically understood by the people who live within them as the result of decisions, policies, historical forces, and the intentions of various actors. These explanations are not wrong, but they are incomplete in a specific way: they attribute the global pattern to intentions operating at the individual or institutional level.
The CA models show that this is systematically misleading. The global pattern is often an emergent property of local interactions, not a direct readout of any individual’s or institution’s intention. Schelling’s segregation need not involve any discriminatory intent. Axelrod’s cultural polarization need not involve any deliberate differentiation strategy. Clarke’s urban sprawl need not involve any planner who chose it. The patterns are the aggregate consequence of millions of individual local decisions, filtered through the CA dynamics of local interaction.
This is not an exculpatory argument — it does not remove the moral weight of individual choices or institutional policies. But it does change what kinds of interventions are effective. If the global pattern is emergent, then changing individual attitudes (below the critical threshold) changes nothing. If a system is in a segregated frozen state, eliminating discriminatory laws may not move it. If a pandemic wave is a CA traveling front, border controls (which are fundamentally spatial interventions) may shift the wave’s path but not eliminate it. Understanding the CA dynamics of the system is a prerequisite for designing effective interventions.
Conway’s Life produced no policy implications. Schelling’s model did. Urban growth models inform zoning regulations. Spatial epidemic models inform public health responses. The Game of Life, it turns out, has practical implications — it just took a few decades and the application to human systems to find them.